diff options
Diffstat (limited to 'llvm/test/CodeGen')
57 files changed, 26750 insertions, 25430 deletions
diff --git a/llvm/test/CodeGen/AArch64/freeze.ll b/llvm/test/CodeGen/AArch64/freeze.ll index 0c56e1b..d428b6a 100644 --- a/llvm/test/CodeGen/AArch64/freeze.ll +++ b/llvm/test/CodeGen/AArch64/freeze.ll @@ -395,3 +395,37 @@ define i64 @freeze_array() { %t1 = add i64 %v1, %v2 ret i64 %t1 } + +define <8 x i16> @freeze_abdu(<8 x i16> %a, <8 x i16> %b) { +; CHECK-SD-LABEL: freeze_abdu: +; CHECK-SD: // %bb.0: +; CHECK-SD-NEXT: uaba v0.8h, v0.8h, v1.8h +; CHECK-SD-NEXT: ret +; +; CHECK-GI-LABEL: freeze_abdu: +; CHECK-GI: // %bb.0: +; CHECK-GI-NEXT: uabd v1.8h, v0.8h, v1.8h +; CHECK-GI-NEXT: add v0.8h, v0.8h, v1.8h +; CHECK-GI-NEXT: ret + %d = call <8 x i16> @llvm.aarch64.neon.uabd.v8i16(<8 x i16> %a, <8 x i16> %b) + %f = freeze <8 x i16> %d + %r = add <8 x i16> %a, %f + ret <8 x i16> %r +} + +define <8 x i16> @freeze_abds(<8 x i16> %a, <8 x i16> %b) { +; CHECK-SD-LABEL: freeze_abds: +; CHECK-SD: // %bb.0: +; CHECK-SD-NEXT: saba v0.8h, v0.8h, v1.8h +; CHECK-SD-NEXT: ret +; +; CHECK-GI-LABEL: freeze_abds: +; CHECK-GI: // %bb.0: +; CHECK-GI-NEXT: sabd v1.8h, v0.8h, v1.8h +; CHECK-GI-NEXT: add v0.8h, v0.8h, v1.8h +; CHECK-GI-NEXT: ret + %d = call <8 x i16> @llvm.aarch64.neon.sabd.v8i16(<8 x i16> %a, <8 x i16> %b) + %f = freeze <8 x i16> %d + %r = add <8 x i16> %a, %f + ret <8 x i16> %r +} diff --git a/llvm/test/CodeGen/AArch64/sve-intrinsics-ldst-ext.ll b/llvm/test/CodeGen/AArch64/sve-intrinsics-ldst-ext.ll index 4153f0b..9698f1a 100644 --- a/llvm/test/CodeGen/AArch64/sve-intrinsics-ldst-ext.ll +++ b/llvm/test/CodeGen/AArch64/sve-intrinsics-ldst-ext.ll @@ -231,3 +231,27 @@ define <vscale x 8 x i64> @sload_8i8_8i64(ptr %a) { %aext = sext <vscale x 8 x i8> %aval to <vscale x 8 x i64> ret <vscale x 8 x i64> %aext } + +; Ensure we don't try to promote a predicate load to a sign-extended load. +define <vscale x 16 x i8> @sload_16i1_16i8(ptr %addr) { +; CHECK-LABEL: sload_16i1_16i8: +; CHECK: // %bb.0: +; CHECK-NEXT: ldr p0, [x0] +; CHECK-NEXT: mov z0.b, p0/z, #-1 // =0xffffffffffffffff +; CHECK-NEXT: ret + %load = load <vscale x 16 x i1>, ptr %addr + %zext = sext <vscale x 16 x i1> %load to <vscale x 16 x i8> + ret <vscale x 16 x i8> %zext +} + +; Ensure we don't try to promote a predicate load to a zero-extended load. +define <vscale x 16 x i8> @zload_16i1_16i8(ptr %addr) { +; CHECK-LABEL: zload_16i1_16i8: +; CHECK: // %bb.0: +; CHECK-NEXT: ldr p0, [x0] +; CHECK-NEXT: mov z0.b, p0/z, #1 // =0x1 +; CHECK-NEXT: ret + %load = load <vscale x 16 x i1>, ptr %addr + %zext = zext <vscale x 16 x i1> %load to <vscale x 16 x i8> + ret <vscale x 16 x i8> %zext +} diff --git a/llvm/test/CodeGen/AMDGPU/add.v2i16.ll b/llvm/test/CodeGen/AMDGPU/add.v2i16.ll index 50d20e9..6cb236d 100644 --- a/llvm/test/CodeGen/AMDGPU/add.v2i16.ll +++ b/llvm/test/CodeGen/AMDGPU/add.v2i16.ll @@ -780,7 +780,8 @@ define amdgpu_kernel void @v_test_add_v2i16_zext_to_v2i64(ptr addrspace(1) %out, ; GFX11-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 ; GFX11-TRUE16-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0x3ff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 2, v0 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: global_load_b32 v1, v0, s[2:3] glc dlc @@ -789,11 +790,12 @@ define amdgpu_kernel void @v_test_add_v2i16_zext_to_v2i64(ptr addrspace(1) %out, ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_pk_add_u16 v0, v1, v0 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_and_b32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshl_or_b32 v2, 0, 16, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l +; GFX11-TRUE16-NEXT: v_lshl_or_b32 v2, v2, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v1 ; GFX11-TRUE16-NEXT: global_store_b128 v1, v[0:3], s[0:1] ; GFX11-TRUE16-NEXT: s_endpgm ; diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll index cb2f0f2..0d5f538 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll @@ -6309,64 +6309,64 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) @@ -6394,50 +6394,50 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2] @@ -6498,50 +6498,50 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v1 ; GFX11-TRUE16-NEXT: .LBB12_4: ; %end @@ -6549,319 +6549,314 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, 0 ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v161.l -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v39 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v1.h, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v2.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff, v66 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v55, v39 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v66 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v67 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v151.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v55, v65 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v2.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v66, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v55 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v65 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v66 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v4.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v5.l, v33.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v39, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v64 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v64 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v145.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v55, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v144.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v54 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v135.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v135.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v53 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v55 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v64 -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v53, v55 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v9.l, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v132.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v53, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v11.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v131.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v53, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v129.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v53 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v53 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v118.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v51, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v51, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v116.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v50, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v114.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v50, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v113.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v49, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v112.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v50 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v50 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v49, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v48, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v49 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v48, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v22.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v49 +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v98.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v97.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v39 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v55, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v12, v39 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v16, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v15, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v17, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v20, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v22, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v80.l -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v26, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v21 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v69.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v68.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v65 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v53, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v38, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v35, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v25, v26 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v29, v30 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v33, v24 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v87.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v25.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v27.l, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v28.l, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v80.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v71.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v69.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v30.l, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v39 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v31.l, v31.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v68.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v32.l, v32.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v39 ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v32i32_to_v128i8: @@ -15418,63 +15413,63 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:384 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:372 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:368 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:364 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:360 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:352 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:356 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:348 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:344 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:336 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:328 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:324 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:312 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:316 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:288 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:292 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:288 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:284 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:280 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:272 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:264 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:276 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:272 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:268 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:264 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260 ; GFX11-TRUE16-NEXT: s_clause 0x1f -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:256 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:256 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:248 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:240 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:232 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:228 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:224 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:220 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:216 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:232 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:228 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:224 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:220 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:216 ; GFX11-TRUE16-NEXT: scratch_load_b32 v103, off, s32 offset:388 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:8 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:16 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:24 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:32 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:40 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:16 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:40 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:48 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:56 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:64 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:72 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:80 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:88 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:104 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:112 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:64 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:72 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:80 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:96 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:112 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v161, off, s32 offset:136 @@ -15488,146 +15483,144 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:212 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:204 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:212 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:204 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:196 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:188 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:180 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:172 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:164 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:140 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:132 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:124 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:116 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:108 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:100 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:180 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:172 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:164 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:156 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:148 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:140 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:132 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:84 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:76 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:60 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:52 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v114, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v30.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v28.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v26.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.l, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.l, v6.l +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v116, off, s32 offset:4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.l, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.l, v0.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v23.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v50.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v49.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v55.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v69.l, 8, v69.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(61) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v68.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(59) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v68.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54) ; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v103 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(53) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v80.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(52) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v81.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v82.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v82.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(50) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v82.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v83.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(49) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v83.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v84.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(48) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v86.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.h, 8, v85.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.l, 8, v86.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v86.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(46) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v87.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(45) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v96.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(44) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v96.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.h, 8, v97.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(42) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.l, 8, v99.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v99.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(41) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v99.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(40) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v101.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v101.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(38) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v160.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(37) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.l, 8, v160.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v160.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(36) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v161.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v161.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v161.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v161.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(34) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v162.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v162.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v162.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(32) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v163.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v163.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(30) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.l, 8, v164.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v164.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.l, 8, v164.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(28) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v165.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v165.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v69.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v69.h, 8, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.l, 8, v55.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.h, 8, v54.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v54.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v53.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v52.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v49.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) @@ -15641,720 +15634,746 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB14_3: ; %cmp.false -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v151.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v148.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v148.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v150.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v150.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v151.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v151.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v145.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v144.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v135.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v146.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v132.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v135.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v146.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v147.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v131.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v132.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v147.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v131.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v119.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v133.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v119.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v130.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v133.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v134.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v118.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v128.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v134.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v11, v12 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v114.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v128.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v150.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v146.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v151.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v150.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v147.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v0.h, v149.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v149.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v144.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v133.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v148.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v145.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v1.h, v147.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v129.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v130.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v113.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v116.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v8.h, v129.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v102.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v102.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v101.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v116.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v117.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v8.h, v117.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v118.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v97.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v131.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v2.l, v146.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v131.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v135.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v3.l, v145.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v128.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v118.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v116.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v4.l, v135.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v134.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v5.l, v133.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v132.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v114.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v6.l, v132.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v130.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v7.l, v130.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v102.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v16, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v97.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v96.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v103.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v112.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v85.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v13.h, v112.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v113.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v85.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v84.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v99.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v83.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v99.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v100.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v17, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v13.h, v100.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v19, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v71.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v70.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v86.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v8.l, v128.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v119.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v9.l, v118.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v102.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v117.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v97.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v10.l, v116.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v100.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v115.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v11.l, v114.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v11.h, v113.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v85.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v12.l, v113.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v103.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v21, v22 -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v71.l -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v17.h, v86.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v87.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v96.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v67.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v18.h, v87.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v67.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v66.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v66.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v17.h, v81.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v82.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v18.h, v82.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v22, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v52.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v68.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v69.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v13.l, v103.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v101.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v81.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v14.l, v101.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v14.h, v99.h +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v71.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v15.l, v99.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v15.h, v98.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v16.l, v97.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v71.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v16.h, v96.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v67.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v26, v27 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v22.h, v68.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v23.h, v69.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v70.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v20, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v26 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v38.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v55.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v22.h, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v23.h, v65.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v29, v30 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v53.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v54.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v17.l, v87.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v17.h, v86.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v64.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v18.l, v85.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v67.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v18.h, v84.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v51.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v19.l, v83.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v19.h, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v48.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v20.l, v82.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v20.h, v80.h +; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v21.l, v80.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v21.h, v70.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v37 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v26, v39 -; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v27.h, v53.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v28.h, v54.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v29 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v25, v38 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v31 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v27.h, v50.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v50.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v28.h, v51.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 16, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v31 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v29 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v36 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v37, v38 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v30, v33 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v34 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v22.l, v70.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v22.h, v68.h +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v38.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v23.l, v68.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v38.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v23.h, v66.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v36.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v24.l, v66.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v37.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v24.h, v65.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v36.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v27, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v28 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v25.l, v64.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v25.h, v55.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v28, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v26.l, v55.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v26.h, v54.h +; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v27.l, v54.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v28.h, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v28.l, v53.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v33.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v28.h, v52.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v33.l, v29.h, v51.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v34 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v29.l, v50.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v30.l, v50.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v32.l, v49.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB14_2 ; GFX11-TRUE16-NEXT: .LBB14_4: ; %cmp.true -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v148.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v145.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v151.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v148.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v150.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v146.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v147.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v150.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v150.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v151.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v145.h, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v151.l, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v149.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v149.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v147.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v134.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v144.h, 3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.l, 3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v135.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v135.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v148.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v146.h, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v147.h, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.l, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v2 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v132.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v131.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v131.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v119.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v132.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v133.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v130.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v145.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v145.h, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v131.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v119.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v118.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v133.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v134.h, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v134.l, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v128.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v128.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v11, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v135.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v135.l, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.l, v32.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v134.l, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v118.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v133.l, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v132.h, v6.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v114.h, 3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v114.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v113.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v132.l, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v117.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v114.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v129.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v130.l, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v116.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v116.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v102.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v101.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v102.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.l, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v112.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v128.l, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v8.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v117.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v119.h, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v102.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l ; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v118.l, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v102.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v117.h, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v100.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v116.l, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v100.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v115.h, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v97.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v114.h, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v103.l, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v117.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v97.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v103.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v16, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v97.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v96.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v85.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.l, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v85.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v112.h, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v113.l, v14.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v99.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v84.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v99.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v84.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v83.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v17, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v80.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v87.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v96.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v103.h, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v85.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v103.l, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v100.l, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v80.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v100.h, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v101.h, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v83.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v15.l ; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v101.l, v14.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v84.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v99.h, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v99.l, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v98.l, v16.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v71.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v97.l, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v71.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v86.l, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v19, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v71.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v70.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v86.h, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v21, v22 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v71.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v67.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v87.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v96.l, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v67.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v16.h ; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v87.h, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v81.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v66.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v66.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v52.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v81.h, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v87.l, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v86.h, v18.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v67.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v85.l, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v67.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v22, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v52.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v82.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v82.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.l, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v48.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v21 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v84.h, v19.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.h, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v82.h, v20.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v51.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v82.l, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v52.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v80.h, v21.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v80.l, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v68.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v69.l, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v24, v25 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v68.h, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v69.h, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v26, v27 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v29 -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v39.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v38.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v70.l, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v37.h, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v70.h, v22.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v39.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.l, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v55.h, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v64.h, v23.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v20, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v26 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v64.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v65.l, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v24 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v22.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v68.h, v23.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v68.l, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v29, v30 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v65.h, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v37.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v66.h, v24.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v64.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v26, v31 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v65.l, v25.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v25.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v27, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v55.l, v26.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v55.h, v26.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v35.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v28 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v26.h ; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v27.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v53.l, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v54.l, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v55.l, v29.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v53.h, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v54.h, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v37 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v26, v39 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v28.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v29 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v34.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v33.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v53.h, v28.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v33.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v49.h, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v50.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v50.h, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v51.l, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v51.h, v29.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v25, v38 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v31 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v53.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 16, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v31 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v29 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v36 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v37, v38 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v30, v33 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v34 +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.l, v30.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v33 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v50.l, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.l, 0x300, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v30.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33 +; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v49.h, v30.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v32.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -42137,64 +42156,64 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) @@ -42222,50 +42241,50 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2] @@ -42309,50 +42328,50 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v1 ; GFX11-TRUE16-NEXT: .LBB36_4: ; %end @@ -42360,319 +42379,314 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, 0 ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v161.l -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v39 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v1.h, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v2.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff, v66 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v55, v39 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v66 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v67 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v151.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v55, v65 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v2.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v66, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v55 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v65 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v66 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v4.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v5.l, v33.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v39, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v64 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v64 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v145.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v55, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v144.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v54 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v135.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v135.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v53 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v55 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v64 -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v53, v55 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v9.l, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v132.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v53, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v11.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v131.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v53, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v129.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v53 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v53 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v118.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v51, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v51, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v116.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v50, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v114.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v50, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v113.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v49, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v112.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v50 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v50 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v49, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v48, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v49 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v48, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v22.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v49 +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v98.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v97.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v39 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v55, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v12, v39 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v16, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v15, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v17, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v20, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v22, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v80.l -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v26, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v21 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v69.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v68.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v65 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v53, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v38, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v35, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v25, v26 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v29, v30 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v33, v24 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v87.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v25.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v27.l, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v28.l, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v80.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v71.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v69.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v30.l, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v39 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v31.l, v31.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v68.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v32.l, v32.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v39 ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v32f32_to_v128i8: @@ -52196,63 +52210,63 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:384 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:372 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:368 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:364 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:360 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:352 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:356 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:348 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:344 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:336 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:328 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:324 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:312 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:316 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:288 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:292 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:288 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:284 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:280 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:272 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:264 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:276 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:272 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:268 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:264 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260 ; GFX11-TRUE16-NEXT: s_clause 0x1f -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:256 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:256 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:248 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:240 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:232 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:228 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:224 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:220 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:216 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:232 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:228 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:224 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:220 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:216 ; GFX11-TRUE16-NEXT: scratch_load_b32 v103, off, s32 offset:388 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:8 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:16 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:24 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:32 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:40 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:16 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:40 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:48 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:56 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:64 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:72 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:80 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:88 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:104 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:112 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:64 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:72 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:80 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:96 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:112 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v161, off, s32 offset:136 @@ -52266,146 +52280,144 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:212 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:204 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:212 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:204 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:196 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:188 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:180 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:172 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:164 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:140 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:132 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:124 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:116 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:108 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:100 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:180 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:172 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:164 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:156 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:148 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:140 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:132 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:84 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:76 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:60 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:52 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v114, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v30.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v28.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v26.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.l, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.l, v6.l +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v116, off, s32 offset:4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.l, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.l, v0.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v23.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v50.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v49.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v55.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v69.l, 8, v69.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(61) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v68.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(59) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v68.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54) ; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v103 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(53) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v80.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(52) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v81.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v82.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v82.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(50) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v82.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v83.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(49) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v83.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v84.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(48) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v86.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.h, 8, v85.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.l, 8, v86.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v86.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(46) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v87.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(45) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v96.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(44) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v96.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.h, 8, v97.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(42) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.l, 8, v99.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v99.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(41) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v99.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(40) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v101.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v101.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(38) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v160.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(37) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.l, 8, v160.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v160.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(36) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v161.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v161.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v161.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v161.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(34) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v162.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v162.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v162.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(32) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v163.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v163.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(30) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.l, 8, v164.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v164.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.l, 8, v164.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(28) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v165.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v165.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v69.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v69.h, 8, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.l, 8, v55.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.h, 8, v54.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v54.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v53.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v52.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v49.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) @@ -52419,720 +52431,746 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB38_3: ; %cmp.false -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v151.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v148.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v148.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v150.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v150.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v151.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v151.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v145.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v144.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v135.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v146.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v132.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v135.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v146.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v147.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v131.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v132.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v147.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v131.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v119.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v133.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v119.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v130.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v133.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v134.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v118.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v128.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v134.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v11, v12 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v114.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v128.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v150.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v146.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v151.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v150.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v147.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v0.h, v149.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v149.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v144.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v133.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v148.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v145.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v1.h, v147.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v129.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v130.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v113.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v116.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v8.h, v129.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v102.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v102.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v101.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v116.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v117.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v8.h, v117.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v118.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v97.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v131.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v2.l, v146.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v131.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v135.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v3.l, v145.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v128.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v118.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v116.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v4.l, v135.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v134.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v5.l, v133.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v132.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v114.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v6.l, v132.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v130.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v7.l, v130.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v102.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v16, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v97.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v96.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v103.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v112.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v85.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v13.h, v112.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v113.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v85.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v84.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v99.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v83.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v99.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v100.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v17, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v13.h, v100.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v19, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v71.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v70.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v86.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v8.l, v128.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v119.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v9.l, v118.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v102.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v117.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v97.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v10.l, v116.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v100.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v115.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v11.l, v114.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v11.h, v113.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v85.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v12.l, v113.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v103.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v21, v22 -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v71.l -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v17.h, v86.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v87.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v96.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v67.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v18.h, v87.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v67.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v66.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v66.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v17.h, v81.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v82.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v18.h, v82.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v22, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v52.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v68.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v69.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v13.l, v103.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v101.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v81.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v14.l, v101.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v14.h, v99.h +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v71.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v15.l, v99.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v15.h, v98.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v16.l, v97.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v71.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v16.h, v96.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v67.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v26, v27 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v22.h, v68.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v23.h, v69.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v70.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v20, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v26 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v38.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v55.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v22.h, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v23.h, v65.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v29, v30 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v53.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v54.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v17.l, v87.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v17.h, v86.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v64.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v18.l, v85.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v67.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v18.h, v84.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v51.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v19.l, v83.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v19.h, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v48.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v20.l, v82.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v20.h, v80.h +; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v21.l, v80.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v21.h, v70.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v37 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v26, v39 -; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v27.h, v53.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v28.h, v54.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v29 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v25, v38 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v31 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v27.h, v50.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v50.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v28.h, v51.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 16, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v31 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v29 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v36 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v37, v38 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v30, v33 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v34 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v22.l, v70.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v22.h, v68.h +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v38.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v23.l, v68.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v38.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v23.h, v66.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v36.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v24.l, v66.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v37.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v24.h, v65.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v36.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v27, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v28 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v25.l, v64.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v25.h, v55.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v28, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v26.l, v55.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v26.h, v54.h +; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v27.l, v54.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v28.h, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v28.l, v53.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v33.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v28.h, v52.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v33.l, v29.h, v51.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v34 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v29.l, v50.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v30.l, v50.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v32.l, v49.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB38_2 ; GFX11-TRUE16-NEXT: .LBB38_4: ; %cmp.true -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v148.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v145.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v151.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v148.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v150.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v146.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v147.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v150.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v150.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v151.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v145.h, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v151.l, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v149.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v149.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v147.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v134.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v144.h, 3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.l, 3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v135.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v135.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v148.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v146.h, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v147.h, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.l, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v2 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v132.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v131.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v131.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v119.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v132.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v133.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v130.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v145.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v145.h, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v131.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v119.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v118.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v133.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v134.h, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v134.l, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v128.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v128.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v11, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v135.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v135.l, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.l, v32.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v134.l, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v118.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v133.l, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v132.h, v6.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v114.h, 3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v114.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v113.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v132.l, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v117.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v114.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v129.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v130.l, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v116.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v116.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v102.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v101.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v102.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.l, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v112.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v128.l, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v8.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v117.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v119.h, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v102.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l ; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v118.l, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v102.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v117.h, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v100.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v116.l, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v100.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v115.h, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v97.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v114.h, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v103.l, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v117.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v97.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v103.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v16, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v97.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v96.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v85.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.l, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v85.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v112.h, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v113.l, v14.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v99.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v84.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v99.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v84.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v83.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v17, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v80.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v87.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v96.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v103.h, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v85.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v103.l, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v100.l, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v80.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v100.h, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v101.h, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v83.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v15.l ; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v101.l, v14.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v84.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v99.h, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v99.l, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v98.l, v16.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v71.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v97.l, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v71.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v86.l, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v19, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v71.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v70.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v86.h, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v21, v22 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v71.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v67.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v87.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v96.l, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v67.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v16.h ; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v87.h, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v81.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v66.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v66.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v52.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v81.h, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v87.l, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v86.h, v18.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v67.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v85.l, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v67.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v22, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v52.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v82.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v82.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.l, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v48.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v21 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v84.h, v19.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.h, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v82.h, v20.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v51.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v82.l, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v52.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v80.h, v21.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v80.l, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v68.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v69.l, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v24, v25 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v68.h, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v69.h, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v26, v27 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v29 -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v39.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v38.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v70.l, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v37.h, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v70.h, v22.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v39.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.l, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v55.h, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v64.h, v23.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v20, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v26 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v64.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v65.l, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v24 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v22.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v68.h, v23.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v68.l, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v29, v30 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v65.h, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v37.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v66.h, v24.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v64.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v26, v31 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v65.l, v25.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v25.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v27, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v55.l, v26.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v55.h, v26.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v35.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v28 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v26.h ; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v27.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v53.l, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v54.l, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v55.l, v29.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v53.h, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v54.h, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v37 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v26, v39 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v28.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v29 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v34.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v33.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v53.h, v28.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v33.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v49.h, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v50.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v50.h, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v51.l, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v51.h, v29.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v25, v38 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v31 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v53.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 16, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v31 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v29 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v36 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v37, v38 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v30, v33 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v34 +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.l, v30.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v33 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v50.l, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.l, 0x300, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v30.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33 +; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v49.h, v30.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v32.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -77900,64 +77938,64 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) @@ -77985,50 +78023,50 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2] @@ -78097,50 +78135,50 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v1 ; GFX11-TRUE16-NEXT: .LBB56_4: ; %end @@ -78148,319 +78186,314 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, 0 ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v161.l -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v39 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v1.h, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v2.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff, v66 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v55, v39 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v66 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v67 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v151.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v55, v65 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v2.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v66, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v55 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v65 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v66 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v4.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v5.l, v33.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v39, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v64 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v64 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v145.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v55, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v144.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v54 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v135.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v135.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v53 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v55 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v64 -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v53, v55 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v9.l, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v132.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v53, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v11.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v131.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v53, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v129.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v53 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v53 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v118.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v51, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v51, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v116.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v50, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v114.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v50, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v113.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v49, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v112.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v50 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v50 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v49, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v48, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v49 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v48, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v22.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v49 +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v98.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v97.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v39 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v55, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v12, v39 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v16, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v15, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v17, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v20, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v22, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v80.l -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v26, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v21 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v69.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v68.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v65 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v53, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v38, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v35, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v25, v26 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v29, v30 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v33, v24 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v87.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v25.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v27.l, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v28.l, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v80.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v71.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v69.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v30.l, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v39 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v31.l, v31.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v68.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v32.l, v32.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v39 ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v16i64_to_v128i8: @@ -87027,63 +87060,63 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:384 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:372 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:368 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:364 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:360 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:352 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:356 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:348 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:344 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:336 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:328 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:324 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:312 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:316 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:288 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:292 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:288 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:284 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:280 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:272 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:264 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:276 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:272 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:268 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:264 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260 ; GFX11-TRUE16-NEXT: s_clause 0x1f -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:256 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:256 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:248 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:240 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:232 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:228 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:224 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:220 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:216 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:232 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:228 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:224 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:220 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:216 ; GFX11-TRUE16-NEXT: scratch_load_b32 v103, off, s32 offset:388 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:8 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:16 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:24 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:32 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:40 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:16 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:40 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:48 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:56 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:64 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:72 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:80 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:88 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:104 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:112 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:64 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:72 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:80 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:96 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:112 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v161, off, s32 offset:136 @@ -87097,146 +87130,144 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:212 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:204 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:212 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:204 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:196 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:188 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:180 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:172 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:164 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:140 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:132 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:124 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:116 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:108 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:100 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:180 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:172 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:164 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:156 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:148 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:140 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:132 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:84 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:76 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:60 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:52 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v114, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v30.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v28.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v26.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.l, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.l, v6.l +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v116, off, s32 offset:4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.l, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.l, v0.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v23.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v50.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v49.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v55.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v69.l, 8, v69.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(61) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v68.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(59) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v68.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54) ; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v103 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(53) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v80.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(52) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v81.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v82.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v82.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(50) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v82.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v83.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(49) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v83.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v84.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(48) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v86.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.h, 8, v85.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.l, 8, v86.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v86.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(46) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v87.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(45) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v96.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(44) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v96.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.h, 8, v97.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(42) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.l, 8, v99.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v99.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(41) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v99.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(40) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v101.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v101.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(38) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v160.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(37) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.l, 8, v160.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v160.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(36) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v161.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v161.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v161.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v161.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(34) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v162.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v162.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v162.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(32) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v163.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v163.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(30) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.l, 8, v164.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v164.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.l, 8, v164.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(28) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v165.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v165.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v69.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v69.h, 8, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.l, 8, v55.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.h, 8, v54.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v54.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v53.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v52.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v49.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) @@ -87250,720 +87281,746 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB58_3: ; %cmp.false -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v151.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v148.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v148.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v150.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v150.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v151.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v151.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v145.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v144.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v135.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v146.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v132.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v135.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v146.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v147.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v131.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v132.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v147.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v131.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v119.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v133.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v119.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v130.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v133.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v134.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v118.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v128.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v134.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v11, v12 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v114.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v128.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v150.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v146.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v151.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v150.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v147.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v0.h, v149.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v149.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v144.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v133.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v148.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v145.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v1.h, v147.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v129.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v130.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v113.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v116.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v8.h, v129.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v102.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v102.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v101.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v116.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v117.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v8.h, v117.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v118.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v97.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v131.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v2.l, v146.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v131.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v135.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v3.l, v145.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v128.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v118.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v116.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v4.l, v135.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v134.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v5.l, v133.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v132.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v114.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v6.l, v132.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v130.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v7.l, v130.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v102.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v16, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v97.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v96.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v103.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v112.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v85.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v13.h, v112.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v113.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v85.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v84.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v99.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v83.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v99.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v100.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v17, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v13.h, v100.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v19, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v71.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v70.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v86.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v8.l, v128.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v119.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v9.l, v118.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v102.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v117.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v97.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v10.l, v116.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v100.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v115.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v11.l, v114.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v11.h, v113.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v85.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v12.l, v113.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v103.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v21, v22 -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v71.l -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v17.h, v86.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v87.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v96.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v67.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v18.h, v87.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v67.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v66.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v66.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v17.h, v81.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v82.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v18.h, v82.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v22, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v52.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v68.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v69.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v13.l, v103.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v101.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v81.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v14.l, v101.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v14.h, v99.h +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v71.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v15.l, v99.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v15.h, v98.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v16.l, v97.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v71.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v16.h, v96.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v67.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v26, v27 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v22.h, v68.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v23.h, v69.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v70.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v20, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v26 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v38.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v55.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v22.h, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v23.h, v65.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v29, v30 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v53.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v54.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v17.l, v87.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v17.h, v86.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v64.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v18.l, v85.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v67.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v18.h, v84.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v51.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v19.l, v83.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v19.h, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v48.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v20.l, v82.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v20.h, v80.h +; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v21.l, v80.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v21.h, v70.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v37 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v26, v39 -; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v27.h, v53.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v28.h, v54.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v29 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v25, v38 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v31 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v27.h, v50.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v50.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v28.h, v51.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 16, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v31 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v29 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v36 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v37, v38 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v30, v33 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v34 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v22.l, v70.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v22.h, v68.h +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v38.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v23.l, v68.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v38.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v23.h, v66.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v36.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v24.l, v66.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v37.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v24.h, v65.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v36.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v27, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v28 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v25.l, v64.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v25.h, v55.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v28, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v26.l, v55.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v26.h, v54.h +; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v27.l, v54.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v28.h, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v28.l, v53.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v33.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v28.h, v52.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v33.l, v29.h, v51.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v34 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v29.l, v50.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v30.l, v50.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v32.l, v49.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB58_2 ; GFX11-TRUE16-NEXT: .LBB58_4: ; %cmp.true -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v148.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v145.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v151.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v148.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v150.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v146.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v147.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v150.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v150.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v151.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v145.h, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v151.l, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v149.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v149.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v147.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v134.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v144.h, 3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.l, 3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v135.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v135.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v148.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v146.h, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v147.h, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.l, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v2 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v132.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v131.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v131.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v119.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v132.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v133.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v130.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v145.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v145.h, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v131.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v119.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v118.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v133.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v134.h, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v134.l, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v128.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v128.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v11, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v135.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v135.l, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.l, v32.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v134.l, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v118.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v133.l, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v132.h, v6.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v114.h, 3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v114.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v113.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v132.l, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v117.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v114.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v129.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v130.l, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v116.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v116.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v102.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v101.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v102.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.l, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v112.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v128.l, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v8.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v117.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v119.h, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v102.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l ; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v118.l, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v102.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v117.h, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v100.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v116.l, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v100.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v115.h, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v97.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v114.h, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v103.l, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v117.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v97.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v103.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v16, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v97.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v96.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v85.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.l, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v85.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v112.h, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v113.l, v14.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v99.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v84.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v99.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v84.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v83.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v17, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v80.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v87.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v96.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v103.h, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v85.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v103.l, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v100.l, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v80.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v100.h, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v101.h, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v83.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v15.l ; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v101.l, v14.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v84.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v99.h, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v99.l, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v98.l, v16.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v71.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v97.l, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v71.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v86.l, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v19, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v71.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v70.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v86.h, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v21, v22 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v71.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v67.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v87.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v96.l, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v67.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v16.h ; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v87.h, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v81.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v66.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v66.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v52.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v81.h, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v87.l, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v86.h, v18.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v67.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v85.l, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v67.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v22, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v52.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v82.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v82.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.l, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v48.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v21 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v84.h, v19.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.h, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v82.h, v20.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v51.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v82.l, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v52.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v80.h, v21.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v80.l, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v68.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v69.l, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v24, v25 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v68.h, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v69.h, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v26, v27 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v29 -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v39.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v38.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v70.l, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v37.h, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v70.h, v22.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v39.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.l, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v55.h, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v64.h, v23.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v20, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v26 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v64.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v65.l, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v24 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v22.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v68.h, v23.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v68.l, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v29, v30 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v65.h, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v37.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v66.h, v24.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v64.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v26, v31 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v65.l, v25.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v25.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v27, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v55.l, v26.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v55.h, v26.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v35.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v28 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v26.h ; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v27.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v53.l, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v54.l, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v55.l, v29.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v53.h, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v54.h, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v37 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v26, v39 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v28.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v29 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v34.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v33.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v53.h, v28.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v33.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v49.h, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v50.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v50.h, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v51.l, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v51.h, v29.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v25, v38 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v31 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v53.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 16, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v31 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v29 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v36 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v37, v38 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v30, v33 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v34 +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.l, v30.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v33 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v50.l, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.l, 0x300, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v30.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33 +; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v49.h, v30.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v32.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -111743,64 +111800,64 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) @@ -111828,50 +111885,50 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2] @@ -111915,50 +111972,50 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v1 ; GFX11-TRUE16-NEXT: .LBB72_4: ; %end @@ -111966,319 +112023,314 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, 0 ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v161.l -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v39 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v1.h, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v2.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff, v66 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v55, v39 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v66 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v67 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v151.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v55, v65 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v2.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v66, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v55 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v65 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v66 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v4.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v5.l, v33.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v39, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v64 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v55, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v64 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v145.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v55, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v144.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v54 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v135.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v135.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v53 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v55 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v64 -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v53, v55 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v9.l, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v132.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v53, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v11.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v131.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v53, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v129.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v53 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v53 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v118.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v51, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v51, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v116.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v50, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v114.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v50, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v113.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v49, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v112.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v50 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v50 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v49, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v48, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v49 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v48, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v22.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v49 +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v98.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v97.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v39 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v55, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v12, v39 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v16, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v15, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v17, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v20, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v22, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v80.l -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v26, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v21 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v69.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v68.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v65 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v53, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v38, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v35, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v25, v26 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v29, v30 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v33, v24 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v87.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v25.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v27.l, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v28.l, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v80.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v71.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v69.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v30.l, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v39 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v31.l, v31.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v68.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v39 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v32.l, v32.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v39 ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v16f64_to_v128i8: @@ -121787,63 +121839,63 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:384 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:372 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:368 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:364 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:360 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:352 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:356 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:348 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:344 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:336 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:328 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:324 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:312 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:316 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:288 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:292 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:288 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:284 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:280 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:272 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:264 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:276 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:272 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:268 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:264 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260 ; GFX11-TRUE16-NEXT: s_clause 0x1f -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:256 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:256 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:248 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:240 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:232 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:228 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:224 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:220 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:216 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:232 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:228 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:224 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:220 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:216 ; GFX11-TRUE16-NEXT: scratch_load_b32 v103, off, s32 offset:388 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:8 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:16 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:24 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:32 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:40 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:16 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:40 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:48 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:56 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:64 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:72 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:80 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:88 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:104 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:112 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:64 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:72 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:80 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:96 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:112 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v161, off, s32 offset:136 @@ -121857,146 +121909,144 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:212 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:204 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:212 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:204 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:196 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:188 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:180 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:172 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:164 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:140 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:132 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:124 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:116 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:108 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:100 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:180 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:172 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:164 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:156 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:148 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:140 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:132 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:84 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:76 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:60 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:52 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v114, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v30.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v28.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v26.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.l, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.l, v6.l +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v116, off, s32 offset:4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.l, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.l, v0.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v23.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v50.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v49.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v55.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v69.l, 8, v69.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(61) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v68.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(59) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v68.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54) ; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v103 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(53) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v80.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(52) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v81.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v82.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v82.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(50) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v82.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v83.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(49) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v83.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v84.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(48) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v86.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.h, 8, v85.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.l, 8, v86.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v86.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(46) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v87.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(45) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v96.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(44) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v96.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.h, 8, v97.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(42) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.l, 8, v99.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v99.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(41) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v99.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(40) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v101.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v101.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(38) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v160.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(37) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.l, 8, v160.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v160.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(36) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v161.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v161.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v161.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v161.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(34) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v162.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v162.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v162.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(32) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v163.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v163.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(30) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.l, 8, v164.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v164.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.l, 8, v164.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(28) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v165.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v165.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v69.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v69.h, 8, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.l, 8, v55.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.h, 8, v54.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v54.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v53.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v52.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v49.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) @@ -122010,720 +122060,746 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB74_3: ; %cmp.false -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v151.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v148.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v148.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v150.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v150.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v151.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v151.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v145.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v144.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v135.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v146.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v132.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v135.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v146.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v147.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v131.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v132.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v147.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v131.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v119.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v133.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v119.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v130.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v133.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v134.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v118.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v128.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v134.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v11, v12 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v114.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v128.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v150.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v146.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v151.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v150.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v147.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v0.h, v149.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v149.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v144.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v133.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v148.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v145.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v1.h, v147.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v129.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v130.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v113.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v116.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v8.h, v129.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v102.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v102.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v101.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v116.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v117.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v8.h, v117.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v118.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v97.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v131.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v2.l, v146.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v131.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v135.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v3.l, v145.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v128.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v118.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v116.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v4.l, v135.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v134.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v5.l, v133.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v132.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v114.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v6.l, v132.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v130.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v7.l, v130.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v102.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v16, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v97.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v96.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v103.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v112.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v85.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v13.h, v112.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v113.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v85.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v84.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v99.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v83.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v99.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v100.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v17, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v13.h, v100.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v19, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v71.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v70.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v86.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v8.l, v128.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v119.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v9.l, v118.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v102.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v117.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v97.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v10.l, v116.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v100.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v115.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v11.l, v114.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v11.h, v113.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v85.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v12.l, v113.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v103.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v21, v22 -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v71.l -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v17.h, v86.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v87.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v96.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v67.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v18.h, v87.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v67.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v66.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v66.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v17.h, v81.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v82.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v18.h, v82.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v22, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v52.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v68.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v69.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v13.l, v103.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v101.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v81.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v14.l, v101.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v14.h, v99.h +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v71.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v15.l, v99.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v15.h, v98.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v16.l, v97.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v71.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v16.h, v96.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v67.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v26, v27 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v22.h, v68.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v23.h, v69.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v70.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v20, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v26 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v38.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v55.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v22.h, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v23.h, v65.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v29, v30 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v53.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v54.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v17.l, v87.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v17.h, v86.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v64.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v18.l, v85.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v67.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v18.h, v84.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v51.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v19.l, v83.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v19.h, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v48.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v20.l, v82.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v20.h, v80.h +; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v21.l, v80.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v21.h, v70.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v37 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v26, v39 -; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v27.h, v53.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v28.h, v54.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v29 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v25, v38 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v31 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v27.h, v50.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v50.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v28.h, v51.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 16, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v31 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v29 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v36 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v37, v38 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v30, v33 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v34 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v22.l, v70.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v22.h, v68.h +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v38.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v23.l, v68.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v38.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v23.h, v66.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v36.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v24.l, v66.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v37.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v24.h, v65.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v36.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v27, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v28 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v25.l, v64.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v25.h, v55.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v28, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v26.l, v55.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v26.h, v54.h +; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v27.l, v54.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v28.h, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v28.l, v53.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v33.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v28.h, v52.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v33.l, v29.h, v51.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v34 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v29.l, v50.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v30.l, v50.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v32.l, v49.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB74_2 ; GFX11-TRUE16-NEXT: .LBB74_4: ; %cmp.true -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v148.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v145.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v151.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v148.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v150.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v146.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v147.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v150.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v150.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v151.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v145.h, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v151.l, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v149.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v149.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v147.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v134.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v144.h, 3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.l, 3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v135.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v135.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v148.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v146.h, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v147.h, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.l, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v2 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v132.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v131.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v131.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v119.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v132.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v133.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v130.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v145.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v145.h, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v131.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v119.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v118.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v133.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v134.h, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v134.l, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v128.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v128.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v11, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v135.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v135.l, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.l, v32.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v134.l, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v118.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v133.l, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v132.h, v6.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v114.h, 3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v114.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v113.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v132.l, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v117.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v114.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v129.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v130.l, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v116.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v116.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v102.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v101.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v102.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.l, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v112.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v128.l, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v8.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v117.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v119.h, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v102.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l ; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v118.l, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v102.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v117.h, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v100.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v116.l, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v100.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v115.h, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v97.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v114.h, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v103.l, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v117.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v97.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v103.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v16, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v97.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v96.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v85.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.l, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v85.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v112.h, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v113.l, v14.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v99.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v84.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v99.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v84.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v83.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v17, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v80.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v87.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v96.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v103.h, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v85.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v103.l, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v100.l, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v80.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v100.h, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v101.h, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v83.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v15.l ; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v101.l, v14.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v84.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v99.h, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v99.l, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v98.l, v16.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v71.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v97.l, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v71.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v86.l, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v19, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v71.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v70.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v86.h, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v21, v22 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v71.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v67.h, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v87.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v96.l, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v67.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v16.h ; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v87.h, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v81.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v16, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v66.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v66.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v52.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v81.h, v17.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v87.l, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v86.h, v18.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v67.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v85.l, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v67.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v22, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v52.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v82.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v82.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.l, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v48.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v21 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v84.h, v19.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.h, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v82.h, v20.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v51.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v82.l, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v52.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v80.h, v21.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v80.l, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v68.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v69.l, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v24, v25 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v68.h, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v69.h, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v26, v27 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v21, v29 -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v39.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v38.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v70.l, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v37.h, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v70.h, v22.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v39.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.l, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v55.h, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v64.h, v23.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v20, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v26 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v64.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v65.l, v23.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v24 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v22.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v68.h, v23.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v68.l, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v29, v30 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v65.h, v24.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v37.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v66.h, v24.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v64.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v26, v31 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v65.l, v25.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v25.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v27, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v55.l, v26.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v55.h, v26.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v35.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v28 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v26.h ; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31 +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v27.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v53.l, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v54.l, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v55.l, v29.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v39, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v53.h, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v54.h, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v37 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v26, v39 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v28.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v29 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v34.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v33.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v53.h, v28.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v33.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v49.h, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v50.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v50.h, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v51.l, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v51.h, v29.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v25, v38 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v31 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v53.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v33, 16, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v31 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v29 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v36 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v37, v38 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v30, v33 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v34 +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.l, v30.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v33 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v50.l, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.l, 0x300, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v30.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33 +; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v49.h, v30.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v32.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -160055,116 +160131,116 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_b32 v33, off, s32 offset:8 ; GFX11-TRUE16-NEXT: scratch_load_b32 v32, off, s32 offset:4 ; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr108_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr107_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr106_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr164_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr106_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr95_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr93_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr180_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr94_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr90_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr180_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr164_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr88_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr165_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr46_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr88_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr75_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr47_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr76_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr178_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr63_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr179_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr62_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr74_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr57_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr72_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr178_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr59_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr73_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr58_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr44_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr56_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr42_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr43_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr41_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr42_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr89_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr40_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr43_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr59_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr182_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr60_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr181_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr94_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr61_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr183_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr57_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr167_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr104_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr176_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr166_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr77_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr76_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr104_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr95_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr93_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr92_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr79_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr73_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr72_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr61_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr74_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr62_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr63_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr58_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr47_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr44_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr60_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr46_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr40_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr41_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr183_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr177_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr181_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr182_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr176_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr166_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr177_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) ; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33 @@ -160187,341 +160263,338 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 24, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v40, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v42, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v56, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v57, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v75, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v78, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v59, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 8, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v76, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v75, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v88, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v93, 24, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v95, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v105, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v106, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v91, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v94, 8, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v106, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v105, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v107, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v108, 8, v1 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 24, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 24, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v177, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v41, 8, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v44, 8, v21 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v47, 24, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v61, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 24, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v73, 8, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v177, 8, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v40, 8, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 8, v21 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v46, 24, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v60, 8, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v74, 8, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v79, 8, v17 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18] -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.h, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.h, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v162.h, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v164.h, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v4.h ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v180.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v164.h, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v165.h, v6.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v46.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v178.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v179.h, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v74.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v45.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v42.h, v10.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v165.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v161.h, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v47.h, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v179.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v178.h, v8.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v73.h, v9.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v44.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v41.h, v10.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v89.h, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v59.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v60.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v94.h, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v77.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v76.h, v14.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v104.h, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v91.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v61.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v57.h, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v104.h, v13.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v78.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v77.h, v14.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v95.h, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v93.h, v16.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v92.h, v16.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.h, v19.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.h, v17.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v18.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v84.h, v19.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.h, v20.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.h, v21.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.h, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v84.h, v22.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.h, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.h, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v87.h, v24.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.h, v21.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.h, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.h, v23.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v87.h, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.h, v24.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v101.h, v25.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.h, v26.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v98.h, v26.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v98.h, v26.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.h, v26.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v112.h, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v99.h, v28.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v100.h, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v100.h, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v99.h, v28.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v113.h, v29.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v103.h, v30.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v102.h, v30.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.h, v31.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v114.h, v32.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v115.h, v32.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v115.h, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v114.h, v32.h ; GFX11-TRUE16-NEXT: .LBB90_2: ; %Flow ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB90_4 ; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true -; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, 0x400000, v17 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v18 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_lshlrev_b32 v18, 16, v18 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v20 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v36, v33, 16, 1 ; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v18, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v18 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33 -; GFX11-TRUE16-NEXT: v_add3_u32 v36, v36, v33, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add3_u32 v37, v37, v18, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v55, v37, v39, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v70, v37, v39 :: v_dual_add_f32 v33, 0x40c00000, v33 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v36, v33, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20 -; GFX11-TRUE16-NEXT: v_bfe_u32 v50, v17, 16, 1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v55.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v70, v36, v38, vcc_lo -; GFX11-TRUE16-NEXT: v_dual_add_f32 v34, 0x40c00000, v34 :: v_dual_add_f32 v35, 0x40c00000, v35 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add3_u32 v37, v50, v17, 0x7fff -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 +; GFX11-TRUE16-NEXT: v_add3_u32 v36, v36, v33, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v17 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v70.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v55, v36, v38, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_add_f32 v34, 0x40c00000, v34 :: v_dual_lshlrev_b32 v17, 16, v17 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v17, 0x40c00000, v17 ; GFX11-TRUE16-NEXT: v_bfe_u32 v48, v34, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, 0x400000, v34 -; GFX11-TRUE16-NEXT: v_bfe_u32 v36, v35, 16, 1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v80, v37, v51, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfe_u32 v50, v17, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, 0x400000, v17 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 ; GFX11-TRUE16-NEXT: v_add3_u32 v18, v48, v34, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v35 +; GFX11-TRUE16-NEXT: v_add3_u32 v37, v50, v17, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v71, v37, v51 :: v_dual_lshlrev_b32 v20, 16, v20 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v20, 0x40c00000, v20 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v11 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v71.h +; GFX11-TRUE16-NEXT: v_bfe_u32 v36, v35, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v20 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v80.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v18, v49, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v18, 0xffff, v33, v70 +; GFX11-TRUE16-NEXT: v_bfi_b32 v18, 0xffff, v33, v55 ; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v20, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v35 ; GFX11-TRUE16-NEXT: v_bfi_b32 v17, 0xffff, v34, v17 ; GFX11-TRUE16-NEXT: v_add3_u32 v34, v36, v35, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v20, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 24, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v73, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v79, 8, v17 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19 +; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v20, 0x7fff +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v74, 8, v18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_add_f32 v19, 0x40c00000, v19 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v81, v33, v37, vcc_lo -; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_lshlrev_b32 v19, 16, v19 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v22 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_lshlrev_b32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v36, 16, 1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v71, v34, v38, vcc_lo -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v81.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 ; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v19, 16, 1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v36, 16, 1 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v80, v34, v38, vcc_lo ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v19 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 +; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v19, 0x7fff ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 ; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v36, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v19, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_add_f32 v22, 0x40c00000, v22 :: v_dual_cndmask_b32 v83, v33, v37 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v81.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v79, 8, v17 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v84, v33, v37, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v22, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v83.h +; GFX11-TRUE16-NEXT: v_add_f32_e32 v35, 0x40c00000, v35 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v84.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v20, v38, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v35, 16, 1 ; GFX11-TRUE16-NEXT: v_add3_u32 v20, v33, v22, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v33, 0x400000, v22 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v82, v20, v33 :: v_dual_add_f32 v35, 0x40c00000, v35 -; GFX11-TRUE16-NEXT: v_bfi_b32 v20, 0xffff, v34, v71 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v82.h -; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v35, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v35 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v47, 24, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v20 ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v35, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v33, 0x40c00000, v38 :: v_dual_cndmask_b32 v84, v19, v39 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v83, v20, v33, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35 +; GFX11-TRUE16-NEXT: v_bfi_b32 v20, 0xffff, v34, v80 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v82, v19, v39, vcc_lo ; GFX11-TRUE16-NEXT: v_bfi_b32 v19, 0xffff, v37, v36 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v24 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 -; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v33, 16, 1 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v83.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v46, 24, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v60, 8, v20 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 +; GFX11-TRUE16-NEXT: v_bfi_b32 v22, 0xffff, v22, v82 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 8, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v36 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v21 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v40, 8, v22 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v33, 0x40c00000, v38 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 +; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v33, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v21, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v21 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33 ; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v33, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v36 ; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v21, 0x7fff -; GFX11-TRUE16-NEXT: v_bfi_b32 v22, 0xffff, v22, v84 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v61, 8, v19 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v85, v34, v37, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v86, v34, v37 :: v_dual_and_b32 v37, 0xffff0000, v23 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v24, 0x40c00000, v24 +; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v24, 16, 1 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 ; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v36, 16, 1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v41, 8, v22 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v21, v35, v38, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v24, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v24, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v24 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 ; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v36, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v85.h -; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v24, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff0000, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfi_b32 v21, 0xffff, v35, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v86, v34, v38, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v86.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v87, v34, v38, vcc_lo ; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v26 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v23, 0x40c00000, v23 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_lshlrev_b32 v26, 16, v26 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v87, v33, v39, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v86.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v23, 0x40c00000, v23 :: v_dual_lshlrev_b32 v26, 16, v26 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 +; GFX11-TRUE16-NEXT: v_bfi_b32 v21, 0xffff, v35, v21 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v26, 0x40c00000, v26 ; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v23, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v37, 16, 1 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v85, v33, v39, vcc_lo ; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, 0x400000, v23 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v37 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 8, v21 ; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v23, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v24, v24, v37, 0x7fff -; GFX11-TRUE16-NEXT: v_add_f32_e32 v26, 0x40c00000, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v44, 8, v21 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v96, v34, v36, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v97, v34, v36, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_f32 v34, 0x40c00000, v38 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v97.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v37, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v37 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v34, 0x40c00000, v38 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v96.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v23, v24, v39, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34 +; GFX11-TRUE16-NEXT: v_add3_u32 v24, v24, v37, 0x7fff ; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v34, 16, 1 -; GFX11-TRUE16-NEXT: v_bfi_b32 v24, 0xffff, v33, v87 -; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v26, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v23, v24, v39, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v35, v37, v34, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v26 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfi_b32 v23, 0xffff, v36, v23 -; GFX11-TRUE16-NEXT: v_add3_u32 v35, v37, v34, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25 -; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v26, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v26 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v87.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v23 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 24, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 8, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v97, v33, v37, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v25, 0x40c00000, v25 +; GFX11-TRUE16-NEXT: v_bfi_b32 v24, 0xffff, v33, v85 +; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v26, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 24, v24 +; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v26, 0x7fff ; GFX11-TRUE16-NEXT: v_bfe_u32 v26, v36, 16, 1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v177, 8, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v98, v35, v38, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v177, 8, v24 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v98, v33, v37, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25 +; GFX11-TRUE16-NEXT: v_add3_u32 v26, v26, v36, 0x7fff +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v96, v35, v38 :: v_dual_add_f32 v25, 0x40c00000, v25 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v28 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v98.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v25, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v25 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v28, 0x40c00000, v28 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25 -; GFX11-TRUE16-NEXT: v_add3_u32 v26, v26, v36, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v25, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v101, v33, v37, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v97.h -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v36, v26, v38 :: v_dual_add_f32 v35, 0x40c00000, v35 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v28, 0x40c00000, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v27 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v101.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v35, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v28, 16, 1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v35 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v101, v33, v37, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v28, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 ; GFX11-TRUE16-NEXT: v_add3_u32 v25, v25, v35, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v26, v38, vcc_lo ; GFX11-TRUE16-NEXT: v_add3_u32 v26, v33, v28, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v33, 0x400000, v28 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v99, v26, v33, vcc_lo +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v27 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v27 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v101.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v100, v26, v33, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35 -; GFX11-TRUE16-NEXT: v_bfi_b32 v26, 0xffff, v34, v98 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v100, v25, v39, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v27, 0x40c00000, v27 +; GFX11-TRUE16-NEXT: v_bfi_b32 v26, 0xffff, v34, v96 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v100.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v99, v25, v39, vcc_lo ; GFX11-TRUE16-NEXT: v_bfi_b32 v25, 0xffff, v37, v36 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v30 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v27 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v33, 0x40c00000, v38 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v30 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v99.h -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 24, v26 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_add_f32 v27, 0x40c00000, v27 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v30, 0x40c00000, v30 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfi_b32 v28, 0xffff, v28, v100 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v26 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v36 ; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v27, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v27 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36 +; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v33, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v28 ; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v27, 0x7fff -; GFX11-TRUE16-NEXT: v_add_f32_e32 v33, 0x40c00000, v38 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v25 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v112, v34, v37, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v33, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v30, 0x40c00000, v30 ; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v33, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v112, v34, v37, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33 ; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v36, 16, 1 +; GFX11-TRUE16-NEXT: v_bfi_b32 v28, 0xffff, v28, v99 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v26 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v27, v35, v38, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v36, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff0000, v29 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 16, v29 @@ -160529,21 +160602,22 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v30 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v112.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v28 ; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v30, 0x7fff +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v25 ; GFX11-TRUE16-NEXT: v_bfi_b32 v27, 0xffff, v35, v27 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1) -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v103, v34, v38 :: v_dual_and_b32 v38, 0xffff0000, v32 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v29, 0x40c00000, v29 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v27 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v29, 16, 1 ; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v102, v33, v39 :: v_dual_add_f32 v37, 0x40c00000, v37 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, 0x400000, v29 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v29, 0x7fff ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v103.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v30, v37, 16, 1 @@ -160556,45 +160630,44 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v113.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v29, v30, v39, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v34, 16, 1 ; GFX11-TRUE16-NEXT: v_bfi_b32 v30, 0xffff, v33, v102 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34 -; GFX11-TRUE16-NEXT: v_bfi_b32 v29, 0xffff, v36, v29 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v31 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v31 ; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v32, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32 +; GFX11-TRUE16-NEXT: v_bfi_b32 v29, 0xffff, v36, v29 ; GFX11-TRUE16-NEXT: v_add3_u32 v35, v37, v34, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v32 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_add_f32 v31, 0x40c00000, v31 ; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v32, 0x7fff -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 24, v30 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v32, v36, 16, 1 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v31 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v29 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v114, v33, v37, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v115, v33, v37, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_lshlrev_b32 v31, 16, v31 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v29 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v115.h +; GFX11-TRUE16-NEXT: v_add_f32_e32 v31, 0x40c00000, v31 +; GFX11-TRUE16-NEXT: v_bfe_u32 v32, v36, 16, 1 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v114, v35, v38, vcc_lo +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v31, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v31 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 ; GFX11-TRUE16-NEXT: v_add3_u32 v32, v32, v36, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v115, v35, v38, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v31, 0x7fff -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v114.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v116, v33, v37, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v116, v33, v37 :: v_dual_and_b32 v35, 0xffff0000, v2 ; GFX11-TRUE16-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v2, 16, v2 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v116.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfe_u32 v31, v35, 16, 1 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v32, v38, vcc_lo ; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v1 @@ -160607,10 +160680,10 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add3_u32 v32, v33, v2, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v33, 0x400000, v2 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v131, v32, v33, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v133, v32, v33, vcc_lo ; GFX11-TRUE16-NEXT: v_add_f32_e32 v33, 0x40c00000, v38 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35 -; GFX11-TRUE16-NEXT: v_bfi_b32 v32, 0xffff, v34, v115 +; GFX11-TRUE16-NEXT: v_bfi_b32 v32, 0xffff, v34, v114 ; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v1, 16, 1 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v132, v31, v39, vcc_lo ; GFX11-TRUE16-NEXT: v_bfi_b32 v31, 0xffff, v37, v36 @@ -160622,9 +160695,9 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33 ; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v33, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v131.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v133.h ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 24, v32 -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v144, v34, v37 :: v_dual_and_b32 v37, 0xffff0000, v3 +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v146, v34, v37 :: v_dual_and_b32 v37, 0xffff0000, v3 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 @@ -160640,252 +160713,240 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v36, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v36 ; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v37, 16, 1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v147, v34, v38, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v148, v34, v38, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v3, 16, 1 ; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v36, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v35.l, v144.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v37, 0x7fff -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v106, 24, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v107, 8, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v149, v33, v39, vcc_lo -; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v37 +; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v3, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v144, v33, v39, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v35, v1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 -; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v108, 8, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v162, v34, v36, vcc_lo +; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v37 +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v33.l, v148.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v35.l, v146.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v105, 24, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v164, v34, v36, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v36.l, v162.h +; GFX11-TRUE16-NEXT: v_add_f32_e32 v34, 0x40c00000, v38 +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v35, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v107, 8, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v36.l, v164.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v4, v39, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v34, 16, 1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v7 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34 ; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v36, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v5 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v33.l, v147.h -; GFX11-TRUE16-NEXT: v_dual_add_f32 v34, 0x40c00000, v38 :: v_dual_lshlrev_b32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v105, 8, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v33, v149 -; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v6, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 -; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v34, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v93, 24, v4 -; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v6, 0x7fff -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v95, 8, v4 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 +; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v33, v144 ; GFX11-TRUE16-NEXT: v_add3_u32 v35, v37, v34, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_add_f32 v5, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v7, 0x40c00000, v7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v94, 8, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v106, 8, v3 +; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v6, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v164, v33, v37, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v108, 8, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v32 +; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v36, 16, 1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v31 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v165, v33, v37, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36 ; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v5, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v165, v35, v38, vcc_lo -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v8 -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v36, 16, 1 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v36, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v34.l, v165.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v161, v35, v38, vcc_lo ; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v5, 0x7fff ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v8 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v35, 0x40c00000, v35 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v36, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v34.l, v164.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v180, v33, v37, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v180, v33, v37 :: v_dual_add_f32 v35, 0x40c00000, v35 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v35, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v37.l, v180.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v6, v38, vcc_lo -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v35, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v35 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 +; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v35, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v8, 16, 1 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v6, v38, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v35 +; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v35, 0x7fff ; GFX11-TRUE16-NEXT: v_add3_u32 v6, v33, v8, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v33, 0x400000, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v178, v6, v33, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v179, v6, v33, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v33, 0x40c00000, v39 ; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v7, 16, 1 -; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v34, v165 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v179, v5, v38, vcc_lo +; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v34, v161 +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v179.h +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v178, v5, v38 :: v_dual_add_f32 v33, 0x40c00000, v39 ; GFX11-TRUE16-NEXT: v_bfi_b32 v5, 0xffff, v37, v36 -; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v33, 16, 1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v9 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v10 ; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v7, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v33, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v7 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v33, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v178.h -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v78, 24, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v46, v35, v37, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36 -; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v179 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v88, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 8, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v33, v34, v38, vcc_lo +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33 +; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v33, 0x7fff +; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v178 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v47, v35, v37, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v36, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v36 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 8, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v59, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 8, v8 ; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v36, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v45, v7, v37, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v33, v34, v38, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v47.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v75, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v88, 8, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v91, 8, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v44, v7, v37, vcc_lo ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v39 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v46.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfe_u32 v36, v9, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, 0x400000, v7 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v36, v36, v9, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 +; GFX11-TRUE16-NEXT: v_add3_u32 v36, v36, v9, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v10, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v10 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v10, 0x7fff -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v42, v35, v38, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v41, v35, v38 :: v_dual_lshlrev_b32 v10, 16, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v37, 0x40c00000, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v45.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v44.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v38, v37, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v35, v42 -; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v7, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v50, 0x400000, v37 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v35, v41 +; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v7, 16, 1 ; GFX11-TRUE16-NEXT: v_add3_u32 v38, v38, v37, 0x7fff -; GFX11-TRUE16-NEXT: v_bfe_u32 v48, v12, 16, 1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v37, 0x40c00000, v51 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v42, 24, v10 ; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v7, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v52, 0x400000, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 24, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v59, v38, v50, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v61, v38, v50 :: v_dual_add_f32 v12, 0x40c00000, v12 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_add3_u32 v48, v48, v12, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v14 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v59.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v74, v35, v49, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v61.h +; GFX11-TRUE16-NEXT: v_bfe_u32 v48, v12, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v52, 0x400000, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v73, v35, v49, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v11 +; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v37, 16, 1 +; GFX11-TRUE16-NEXT: v_add3_u32 v48, v48, v12, 0x7fff ; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_lshlrev_b32 v11, 16, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v56, 8, v10 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v60, v48, v52 :: v_dual_add_f32 v37, 0x40c00000, v51 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v57, v48, v52, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 ; GFX11-TRUE16-NEXT: v_bfe_u32 v49, v14, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v7, v60 -; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v37, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v7, v57 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v36, v39, vcc_lo ; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v11 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v74.h ; GFX11-TRUE16-NEXT: v_add3_u32 v11, v35, v37, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, 0x400000, v37 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v73.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v39, v7, 16, 1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 24, v12 -; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v36, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 8, v12 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v11, v11, v35, vcc_lo ; GFX11-TRUE16-NEXT: v_add_f32_e32 v35, 0x40c00000, v38 ; GFX11-TRUE16-NEXT: v_add3_u32 v37, v39, v7, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v7 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v13 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 ; GFX11-TRUE16-NEXT: v_bfe_u32 v48, v35, 16, 1 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10] +; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v36, v9 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v39 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v89, v37, v38, vcc_lo -; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v35 ; GFX11-TRUE16-NEXT: v_add3_u32 v37, v48, v35, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v35 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, 0x400000, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v57, 8, v9 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v76, v37, v38 :: v_dual_and_b32 v37, 0xffff0000, v16 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v39 :: v_dual_lshlrev_b32 v16, 16, v16 ; GFX11-TRUE16-NEXT: v_add3_u32 v39, v49, v14, 0x7fff -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, 0x400000, v14 ; GFX11-TRUE16-NEXT: v_bfe_u32 v49, v7, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v77, v37, v38, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, 0x400000, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v77, v39, v48, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 ; GFX11-TRUE16-NEXT: v_add3_u32 v14, v49, v7, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, 0x400000, v16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v77.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff0000, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v78, v39, v48, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_bfe_u32 v39, v13, 16, 1 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10] +; GFX11-TRUE16-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v78.h ; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v7, v14, v35 :: v_dual_add_f32 v14, 0x40c00000, v37 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v48, v14, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v37, 0x40c00000, v37 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, 0x400000, v37 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v50, v15, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v39, v13, 16, 1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add3_u32 v35, v39, v13, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v13 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13 ; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v16, 16, 1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v94, v35, v39, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v39, v37, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v37, 0x40c00000, v37 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, 0x400000, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v104, v35, v39, vcc_lo ; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v16, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v39, v37, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 +; GFX11-TRUE16-NEXT: v_bfe_u32 v48, v14, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, 0x400000, v37 +; GFX11-TRUE16-NEXT: v_add3_u32 v39, v39, v37, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v93, v13, v49, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37 ; GFX11-TRUE16-NEXT: v_add3_u32 v35, v48, v14, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, 0x400000, v14 -; GFX11-TRUE16-NEXT: v_add3_u32 v39, v39, v37, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v50, v15, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v91, v13, v49, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37 -; GFX11-TRUE16-NEXT: v_add3_u32 v13, v50, v15, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v94.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v104, v39, v51, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v95, v39, v51, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v104.h +; GFX11-TRUE16-NEXT: v_add3_u32 v13, v50, v15, 0x7fff ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v89.h -; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v38, v76 +; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v38, v77 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v92, v35, v48, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v104.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v91.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v95.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v93.h ; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v39, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v14 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v13, v13, v16, vcc_lo -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v14 ; GFX11-TRUE16-NEXT: v_bfi_b32 v16, 0xffff, v35, v92 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18] @@ -160905,332 +160966,327 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v40, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v75, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v76, 8, v7 ; GFX11-TRUE16-NEXT: .LBB90_4: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v144.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v108.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v68.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v132.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v106.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v133.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v107.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, 0 ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v1.h, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v131.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v107.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v132.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v2.h, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v106.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v78.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v2.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v105.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v162.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v164.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v1.h, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v105.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v2.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v94.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v91.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v4.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v148.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v67.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v2.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v147.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v95.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v93.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v5.l, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v180.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v90.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v66.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v164.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v180.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v90.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v144.h ; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v88.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v165.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v10 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v14 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v66.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v4.l, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v165.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v88.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v5.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v47.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v76.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v58.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v14 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v75.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v161.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v179.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v72.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v6.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v7.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v65.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v46.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v75.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v178.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v63.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v179.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v62.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v14 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v74.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v57.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v73.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v178.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v59.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v14 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v56.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v8.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v44.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v45.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v56.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v18 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v43.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v89.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v41.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v42.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v16, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v10.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v89.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v40.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v14 +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v61.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v183.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v16, v14 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v42.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v43.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v60.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v181.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v11.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v104.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v176.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v166.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v16, v14 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v167.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v57.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v78.h ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v94.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v167.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v59.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v182.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v77.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v163.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v20, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v76.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v104.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v151.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v12, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v91.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v92.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v135.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v55.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v73.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v80.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v79.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v70.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v72.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v16, v18 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v83.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v61.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v12.l, v14.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v16, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v160.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v77.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v95.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v150.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v64, v18, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v93.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v149.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v65, v18, v14 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v79.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v92.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v134.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v13.h, v15.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v66, v18, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v70.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v74.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v67, v18, v14 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v46.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v84.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v63.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v14 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v55.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v62.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v81.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v60.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v13.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v71.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v47.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v16, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v81.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v58.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v13.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v20, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v86.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v17.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v45.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v19, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v13.h, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v83.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v40.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v18.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v19, v14 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v38.l ; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v85.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v44.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v18, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v15, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v96.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v177.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v97.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v182.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v19.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v13.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v82.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v41.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v84.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v183.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v17, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v86.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v176.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v181.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v13.h, v19.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v24 +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v87.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v177.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v22, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v147.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v13.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v101.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v163.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v22, v14 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v85.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v162.l ; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v87.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v166.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v101.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v161.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v97.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v150.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v98.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v148.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v112.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v145.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v20, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v22, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v99.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v100.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v113.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v103.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v26, v27 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v28, v19 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v21 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v102.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v116.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v98.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v36.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.h, v22.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v26, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v28 +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v112.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v23.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v145.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v25, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v26 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.h, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v100.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v135.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v24.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v25, v14 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v113.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v131.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v26 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v25.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v99.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.l, 8, v130.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v13.h, v25.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v28, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v30 +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v103.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v26.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v129.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v28, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v30 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v13.h, v27.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v116.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v128.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v27.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v28, v14 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v102.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.l, 8, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v29 +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v115.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v118.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v13.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v114.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v115.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v25, v26 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v29, v30 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v33, v24 +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v13.h, v28.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v32, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v34 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v114.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v30 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v31, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v30, v14 ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[64:67], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[15:18], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[19:22], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[23:26], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[27:30], off offset:112 ; GFX11-TRUE16-NEXT: s_clause 0x1f ; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:16 @@ -185249,64 +185305,64 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) @@ -185331,52 +185387,52 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20] @@ -185436,371 +185492,366 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v17 ; GFX11-TRUE16-NEXT: .LBB94_4: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v163.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v68.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, 0 ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v161.l -; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v1.h, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v2.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v69.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v68, 0xffff, v68 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v54, v51 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v68 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v67, 16, v69 -; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v2.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v151.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v67 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v2.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v68, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v66.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v51, v54 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v67 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v68 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v67 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v4.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v5.l, v33.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v51, v54 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v65.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v66 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v67 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v51, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v66 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v135.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v65 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v51, v54 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v131.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v65 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v66 -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v65 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v117.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v118.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v11.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v51, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v116.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v64 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v51 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v97.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v49, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v49, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v84.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v48, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v145.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v134.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v128.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v22.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v114.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v113.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v65, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v12, v51 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v145.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v144.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v16, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v132.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v15, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v17, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v20, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v22, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v26, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v21 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v80.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v71.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v55.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 16, v67 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v54, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v38, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v35, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v25, v26 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v29, v30 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v33, v24 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v51 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v25.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v86.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v27.l, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v28.l, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v81.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v71.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v30.l, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v51 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v31.l, v31.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v55.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v32.l, v32.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v51 ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v64f16_to_v128i8: @@ -208007,64 +208058,64 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) @@ -208089,52 +208140,52 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20] @@ -208194,371 +208245,366 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v15 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v28 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v22 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v17 ; GFX11-TRUE16-NEXT: .LBB98_4: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v163.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v68.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, 0 ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v161.l -; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v1.h, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v2.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v69.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v68, 0xffff, v68 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v54, v51 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v68 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v67, 16, v69 -; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v2.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v151.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v67 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v2.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v68, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v66.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v51, v54 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v67 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v68 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v67 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v4.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v5.l, v33.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v51, v54 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.h, v33.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v65.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v66 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v67 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v51, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v66 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v135.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v65 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v51, v54 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v131.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v8.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v65 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v66 -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v65 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v117.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v65 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v118.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v11.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v51, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v116.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v54, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v64 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v11.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v54 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v51 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v97.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v49, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v14.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v49, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v52 +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v84.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v39 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v48, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v145.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v17.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v134.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v128.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v20.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v22.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48 +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v114.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v113.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v39 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v23.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v51.l, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v65, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v12, v51 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v145.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v144.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v16, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v132.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v15, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v17, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v17 -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v20, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v22, v23 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v26, v27 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v21 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v80.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v71.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v55.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 16, v67 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v54, v64 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v38, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v35, v19 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v25, v26 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v27, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v29, v30 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v32 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v33, v24 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v51 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v25.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v86.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v26.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37 +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v27.l, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v28.l, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v81.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36 +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v29.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v51 +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v71.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v30.l, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v51 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v31.l, v31.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l +; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v55.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v32.l, v32.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v51 ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v64i16_to_v128i8: diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll index 34d7ed9..3e96ab1 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll @@ -2675,79 +2675,76 @@ define <4 x i32> @bitcast_v8bf16_to_v4i32(<8 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB22_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v8, v10 :: v_dual_add_f32 v4, 0x40c00000, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v5 :: v_dual_add_f32 v7, 0x40c00000, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v2.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v6, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v1.l +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v11, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v8, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v11 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h +; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v4, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v4 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v4, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v7, v9 :: v_dual_and_b32 v5, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v0 +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v13, v11, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v7 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v4, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v3, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v5 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v7, 16, 1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v8, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v10, v1, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v13, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v11, v12, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v8, 0x7fff +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v7, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v6, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v9 -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v6, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v9, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v8, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v8 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v9, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: v_add3_u32 v9, v13, v0, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v8, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v10, v14, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v7, v11, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v10, v14, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v5.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v7, v3, v11, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v8.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v6 +; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v6, v2 +; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v4, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v8, v7 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v7, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v5, v0 ; GFX11-TRUE16-NEXT: .LBB22_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -4122,18 +4119,18 @@ define <4 x i32> @bitcast_v16i8_to_v4i32(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -4147,107 +4144,103 @@ define <4 x i32> @bitcast_v16i8_to_v4i32(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v10 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2 ; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v7.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v8.h, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v4.h, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v5.h, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -7140,79 +7133,76 @@ define <4 x float> @bitcast_v8bf16_to_v4f32(<8 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB46_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v8, v10 :: v_dual_add_f32 v4, 0x40c00000, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v5 :: v_dual_add_f32 v7, 0x40c00000, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v2.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v6, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v1.l +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v11, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v8, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v11 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h +; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v4, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v4 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v4, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v7, v9 :: v_dual_and_b32 v5, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v0 +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v13, v11, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v7 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v4, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v3, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v5 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v7, 16, 1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v8, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v10, v1, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v13, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v11, v12, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v8, 0x7fff +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v7, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v6, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v9 -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v6, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v9, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v8, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v8 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v9, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: v_add3_u32 v9, v13, v0, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v8, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v10, v14, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v7, v11, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v10, v14, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v5.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v7, v3, v11, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v8.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v6 +; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v6, v2 +; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v4, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v8, v7 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v7, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v5, v0 ; GFX11-TRUE16-NEXT: .LBB46_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -8603,18 +8593,18 @@ define <4 x float> @bitcast_v16i8_to_v4f32(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -8628,107 +8618,103 @@ define <4 x float> @bitcast_v16i8_to_v4f32(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v10 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2 ; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v7.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v8.h, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v4.h, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v5.h, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -11253,79 +11239,76 @@ define <2 x i64> @bitcast_v8bf16_to_v2i64(<8 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB66_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v8, v10 :: v_dual_add_f32 v4, 0x40c00000, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v5 :: v_dual_add_f32 v7, 0x40c00000, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v2.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v6, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v1.l +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v11, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v8, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v11 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h +; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v4, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v4 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v4, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v7, v9 :: v_dual_and_b32 v5, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v0 +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v13, v11, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v7 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v4, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v3, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v5 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v7, 16, 1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v8, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v10, v1, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v13, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v11, v12, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v8, 0x7fff +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v7, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v6, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v9 -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v6, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v9, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v8, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v8 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v9, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: v_add3_u32 v9, v13, v0, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v8, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v10, v14, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v7, v11, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v10, v14, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v5.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v7, v3, v11, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v8.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v6 +; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v6, v2 +; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v4, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v8, v7 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v7, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v5, v0 ; GFX11-TRUE16-NEXT: .LBB66_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -12700,18 +12683,18 @@ define <2 x i64> @bitcast_v16i8_to_v2i64(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -12725,107 +12708,103 @@ define <2 x i64> @bitcast_v16i8_to_v2i64(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v10 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2 ; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v7.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v8.h, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v4.h, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v5.h, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -14952,79 +14931,76 @@ define <2 x double> @bitcast_v8bf16_to_v2f64(<8 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB82_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v8, v10 :: v_dual_add_f32 v4, 0x40c00000, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v5 :: v_dual_add_f32 v7, 0x40c00000, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v2.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v6, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v1.l +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v11, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v8, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v11 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h +; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v4, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v4 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v4, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v7, v9 :: v_dual_and_b32 v5, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v0 +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v13, v11, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v7 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v4, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v3, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v5 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v7, 16, 1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v8, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v10, v1, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v13, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v11, v12, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v8, 0x7fff +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v7, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v6, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v9 -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v6, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v9, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v8, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v8 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v9, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: v_add3_u32 v9, v13, v0, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v8, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v10, v14, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v7, v11, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v10, v14, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v5.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v7, v3, v11, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v8.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v6 +; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v6, v2 +; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v4, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v8, v7 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v7, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v5, v0 ; GFX11-TRUE16-NEXT: .LBB82_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -16407,18 +16383,18 @@ define <2 x double> @bitcast_v16i8_to_v2f64(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -16432,107 +16408,103 @@ define <2 x double> @bitcast_v16i8_to_v2f64(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v10 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2 ; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v7.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v8.h, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v4.h, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v5.h, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -18254,83 +18226,83 @@ define <8 x i16> @bitcast_v8bf16_to_v8i16(<8 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB94_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_add_f32 v4, 0x40c00000, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v5, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v5 -; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v4, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v7, v8 :: v_dual_and_b32 v1, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v13, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_add_f32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v0, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v8, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v3, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v1.l +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v6, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v7, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v7 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v5, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: v_add3_u32 v12, v12, v7, 0x7fff +; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v11, v1, 0x7fff +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, 16, v5 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v10, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v2 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v11, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v12, v13, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v7, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v7 ; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v11, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v7, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v4, 16, 1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v6, 16, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v7, v9, v11, vcc_lo ; GFX11-TRUE16-NEXT: v_add3_u32 v9, v12, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v11, v13, v8, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v8 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX11-TRUE16-NEXT: v_add3_u32 v11, v13, v4, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v11, v12 :: v_dual_and_b32 v3, 0xffff0000, v3 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_add3_u32 v13, v14, v3, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v11, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v13, v14, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v3.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v9, v15, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, 16, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v7, v10, vcc_lo +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, 16, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v8, v10, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.h -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 16, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 16, v7 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.h ; GFX11-TRUE16-NEXT: v_lshl_or_b32 v3, v0, 16, v3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.h ; GFX11-TRUE16-NEXT: v_lshl_or_b32 v2, v1, 16, v2 -; GFX11-TRUE16-NEXT: v_lshl_or_b32 v0, v7, 16, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshl_or_b32 v1, v6, 16, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_lshl_or_b32 v0, v7, 16, v5 +; GFX11-TRUE16-NEXT: v_lshl_or_b32 v1, v4, 16, v6 ; GFX11-TRUE16-NEXT: .LBB94_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -19840,18 +19812,18 @@ define <8 x i16> @bitcast_v16i8_to_v8i16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -19865,107 +19837,103 @@ define <8 x i16> @bitcast_v16i8_to_v8i16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB98_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v10 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_2 ; GFX11-TRUE16-NEXT: .LBB98_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v7.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v8.h, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v4.h, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v5.h, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -21172,79 +21140,79 @@ define <8 x half> @bitcast_v8bf16_to_v8f16(<8 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB102_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v1.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v6, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v2.l +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v6, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v4, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v7, v8 :: v_dual_and_b32 v5, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v3 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v5 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v4, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v1, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v4, 0x7fff ; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v8 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v10, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_cndmask_b32 v1, v7, v9 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v1, v7 :: v_dual_and_b32 v7, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v11, 16, 1 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v8, v10, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v11, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v11 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v7 :: v_dual_add_f32 v7, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v12, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v6, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v7 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_cndmask_b32 v2, v2, v8 +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v7, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v7, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v5, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v7, v8, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v3, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v10, v6, 0x7fff -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v8, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v13, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v8, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v10, v14, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v7.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v10, v14, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v13, v3, 0x7fff ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v11, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v7, v3 -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v5.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v1.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v2.h +; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v7, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v5, v3 +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v8, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v9, v4 ; GFX11-TRUE16-NEXT: .LBB102_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -22758,18 +22726,18 @@ define <8 x half> @bitcast_v16i8_to_v8f16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -22783,107 +22751,103 @@ define <8 x half> @bitcast_v16i8_to_v8f16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB106_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v10 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB106_2 ; GFX11-TRUE16-NEXT: .LBB106_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v7.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v8.h, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v4.h, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v5.h, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -23876,87 +23840,92 @@ define <16 x i8> @bitcast_v8bf16_to_v16i8(<8 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB108_4 ; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v2.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v6, v8 :: v_dual_and_b32 v3, 0xffff0000, v2 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v1, 16, v3 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_add_f32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v14, v4, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v1, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v6, v8 :: v_dual_and_b32 v3, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_lshlrev_b32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v14, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff +; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v9, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v11.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v2, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v13, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v11 -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v9, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v13, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v11 -; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v12, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v7, v9, vcc_lo -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v10 -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v8, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v12, 0x7fff -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 -; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v9, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v8, 0x7fff -; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v3, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v2, v13, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v11, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v5 -; GFX11-TRUE16-NEXT: v_add3_u32 v2, v14, v9, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v10, v15, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add3_u32 v0, v9, v13, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v7, v8, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v13 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13 +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v0, v0, v8 :: v_dual_and_b32 v5, 0xffff0000, v11 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v12.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v7, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v5, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v13, vcc_lo -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, 8, v3 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v10, 0x40c00000, v10 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v8, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v1, 0x7fff +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v10, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v12.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v11, v15, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v5, v14, v10, 0x7fff +; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v2, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v9, v13, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v1, v14 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v5, v16, vcc_lo +; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v13, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v7, 24, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v15, 24, v11 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v7, v2 -; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v9, v1 +; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v9, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v13, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v7, 24, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[16:17], 24, v[10:11] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v2 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[16:17], 24, v[10:11] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[17:18], 24, v[2:3] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v9, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v2 ; GFX11-TRUE16-NEXT: .LBB108_4: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v17.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h @@ -24976,18 +24945,18 @@ define <8 x bfloat> @bitcast_v16i8_to_v8bf16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -25001,107 +24970,103 @@ define <8 x bfloat> @bitcast_v16i8_to_v8bf16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB110_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v8.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v10 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB110_2 ; GFX11-TRUE16-NEXT: .LBB110_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v7.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v8.h, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v4.h, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v5.h, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v8, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll index 2c78e34..5344095 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.16bit.ll @@ -659,7 +659,8 @@ define i16 @bitcast_bf16_to_i16(bfloat %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB6_4 ; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 ; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 @@ -1132,7 +1133,8 @@ define half @bitcast_bf16_to_f16(bfloat %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB10_4 ; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 ; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll index a19567b..f8ffaa4 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll @@ -6296,33 +6296,32 @@ define <8 x i32> @bitcast_v32i8_to_v8i32(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v22.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB26_3 @@ -6334,188 +6333,194 @@ define <8 x i32> @bitcast_v32i8_to_v8i32(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v16.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v15.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v4.h, v12.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v4.h, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v0.h, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v1.h, v14.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v2.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v3.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v12, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v4.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v5.l, v9.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v9.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v6.l, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v7.l, v8.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v19 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2 ; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v14.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v18.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v17.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v15.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v15.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v19.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v13.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v16.h, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v15.h, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v16.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v14.h, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v13.l, v1.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v10.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v11.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v12.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v12.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v12.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v18.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.h, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v21 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v4.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v22.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v30.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v10.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v21 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v11.l, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v8.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v8.h, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v9.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v9.h, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v21 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v9.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v7 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v10 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.h, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v21 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v21 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v8, v21 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -13330,33 +13335,32 @@ define <8 x float> @bitcast_v32i8_to_v8f32(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v22.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB50_3 @@ -13368,188 +13372,194 @@ define <8 x float> @bitcast_v32i8_to_v8f32(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v16.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v15.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v4.h, v12.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v4.h, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v0.h, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v1.h, v14.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v2.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v3.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v12, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v4.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v5.l, v9.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v9.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v6.l, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v7.l, v8.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v19 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2 ; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v14.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v18.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v17.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v15.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v15.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v19.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v13.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v16.h, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v15.h, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v16.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v14.h, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v13.l, v1.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v10.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v11.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v12.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v12.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v12.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v18.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.h, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v21 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v4.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v22.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v30.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v10.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v21 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v11.l, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v8.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v8.h, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v9.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v9.h, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v21 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v9.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v7 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v10 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.h, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v21 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v21 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v8, v21 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -19882,33 +19892,32 @@ define <4 x i64> @bitcast_v32i8_to_v4i64(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v22.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB70_3 @@ -19920,188 +19929,194 @@ define <4 x i64> @bitcast_v32i8_to_v4i64(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v16.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v15.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v4.h, v12.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v4.h, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v0.h, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v1.h, v14.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v2.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v3.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v12, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v4.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v5.l, v9.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v9.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v6.l, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v7.l, v8.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v19 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2 ; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v14.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v18.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v17.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v15.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v15.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v19.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v13.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v16.h, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v15.h, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v16.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v14.h, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v13.l, v1.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v10.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v11.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v12.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v12.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v12.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v18.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.h, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v21 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v4.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v22.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v30.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v10.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v21 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v11.l, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v8.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v8.h, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v9.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v9.h, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v21 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v9.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v7 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v10 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.h, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v21 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v21 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v8, v21 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -25924,33 +25939,32 @@ define <4 x double> @bitcast_v32i8_to_v4f64(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v22.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB86_3 @@ -25962,188 +25976,194 @@ define <4 x double> @bitcast_v32i8_to_v4f64(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v16.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v15.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v4.h, v12.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.h, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v4.h, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v0.h, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v1.h, v14.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v2.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v3.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v12, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v4.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v5.l, v9.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v9.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v6.l, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v7.l, v8.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v19 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2 ; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v14.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v18.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v17.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v15.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v15.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v19.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v13.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v16.h, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v15.h, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v16.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v14.h, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v13.l, v1.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v10.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v11.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v12.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v12.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v21 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v12.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v18.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.h, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v21 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v4.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v22.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v30.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v10.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v21 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v11.l, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v8.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v8.h, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v9.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v9.h, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v21 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v9.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v7 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v6, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v10 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.h, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v21 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v21 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v8, v21 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll index e773b54..0cefbc1 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll @@ -2966,20 +2966,20 @@ define <40 x i8> @bitcast_v10i32_to_v40i8(<10 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 @@ -2995,17 +2995,17 @@ define <40 x i8> @bitcast_v10i32_to_v40i8(<10 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v17, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v18, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB12_2: ; %Flow @@ -3029,17 +3029,17 @@ define <40 x i8> @bitcast_v10i32_to_v40i8(<10 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v17, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v18, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB12_4: ; %end @@ -3047,100 +3047,105 @@ define <40 x i8> @bitcast_v10i32_to_v40i8(<10 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v2.h, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v1.h, v11.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v29 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v16, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v28, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v14, v15 ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v4.l, v11.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v29, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v7.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v11.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v30, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v26, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v13, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v17, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v19, v10 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v15 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v17.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v11, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v15 ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[10:11], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v10i32_to_v40i8: @@ -5033,50 +5038,48 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:24 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:32 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:20 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:4 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v25.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v23.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v21.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v15.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v28.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v28.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(8) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v33.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v33.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v33.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(6) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v34.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v34.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v35.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v35.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v36 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB14_3 @@ -5091,228 +5094,243 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v26.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v25.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v24.h ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v24.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v5.h, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v5.h, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v17, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v12 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v0.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v21.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v1.h, v19.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v2.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v3.l, v15.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v28.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v4.l, v14.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v14.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v25 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v5.l, v13.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v6.h, v13.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v6.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v7.l, v11.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v31.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v25 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v8.l, v10.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v25 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v9.l, v10.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v25 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB14_2 ; GFX11-TRUE16-NEXT: .LBB14_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v26.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v25.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v25.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v20.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v22.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v23.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v21.h, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v18.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v24.h, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v22.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v24.h, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v19.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v15.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v19.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v16.l, 3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v20.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v15.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v18.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v16.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v17.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v21.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v18.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v21.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v19.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v16.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v17.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v27 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v19.h, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v27 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v15.h, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v27 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v18.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v27 +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v14.h, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v15.l, v4.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v27 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v14.l, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v13.h, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v13.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v14.h, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v13.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v14.l, v6.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v30.l, 3 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v31.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v13.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v12.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v27 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v10.h, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v11.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v11.h, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v27 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v12.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v11.l, v7.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v17, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v12 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v11.h, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v10.h, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v27 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v8.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v27 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v27 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -9933,20 +9951,20 @@ define <40 x i8> @bitcast_v10f32_to_v40i8(<10 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 @@ -9962,17 +9980,17 @@ define <40 x i8> @bitcast_v10f32_to_v40i8(<10 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v17, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v18, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB32_2: ; %Flow @@ -9992,17 +10010,17 @@ define <40 x i8> @bitcast_v10f32_to_v40i8(<10 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[13:14], 24, v[5:6] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[14:15], 24, v[3:4] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB32_4: ; %end @@ -10010,100 +10028,105 @@ define <40 x i8> @bitcast_v10f32_to_v40i8(<10 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v2.h, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v1.h, v11.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v29 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v16, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v28, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v14, v15 ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v4.l, v11.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v29, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v7.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v11.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v30, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v26, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v13, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v17, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v19, v10 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v15 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v17.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v11, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v15 ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[10:11], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v10f32_to_v40i8: @@ -12014,50 +12037,48 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:24 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:32 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:20 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:4 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v25.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v23.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v21.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v15.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v28.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v28.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(8) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v33.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v33.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v33.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(6) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v34.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v34.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v35.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v35.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v36 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB34_3 @@ -12072,228 +12093,243 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v26.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v25.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v24.h ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v24.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v5.h, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v5.h, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v17, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v12 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v0.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v21.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v1.h, v19.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v2.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v3.l, v15.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v28.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v4.l, v14.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v14.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v25 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v5.l, v13.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v6.h, v13.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v6.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v25 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v7.l, v11.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v31.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v25 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v8.l, v10.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v25 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v9.l, v10.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v25 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB34_2 ; GFX11-TRUE16-NEXT: .LBB34_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v26.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v25.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v25.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v20.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v22.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v23.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v21.h, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v18.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v24.h, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v22.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v24.h, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v19.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v15.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v19.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v16.l, 3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v20.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v15.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v18.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v16.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v17.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v21.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v18.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v21.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v19.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v16.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v17.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v27 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v19.h, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v27 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v15.h, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v27 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v18.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v27 +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v14.h, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v15.l, v4.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v27 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v14.l, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v13.h, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v13.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v14.h, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v13.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v14.l, v6.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v30.l, 3 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v31.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v13.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v12.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v27 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v10.h, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v11.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v11.h, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v27 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v12.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v11.l, v7.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v17, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v12 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v11.h, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v10.h, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v27 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v8.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v27 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v27 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -16322,20 +16358,20 @@ define <40 x i8> @bitcast_v20i16_to_v40i8(<20 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 @@ -16351,17 +16387,17 @@ define <40 x i8> @bitcast_v20i16_to_v40i8(<20 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v17, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v18, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB48_2: ; %Flow @@ -16385,17 +16421,17 @@ define <40 x i8> @bitcast_v20i16_to_v40i8(<20 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v17, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v18, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB48_4: ; %end @@ -16403,100 +16439,105 @@ define <40 x i8> @bitcast_v20i16_to_v40i8(<20 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v2.h, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v1.h, v11.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v29 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v16, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v28, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v14, v15 ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v4.l, v11.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v29, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v7.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v11.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v30, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v26, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v13, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v17, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v19, v10 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v15 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v17.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v11, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v15 ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[10:11], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v20i16_to_v40i8: @@ -22438,20 +22479,20 @@ define <40 x i8> @bitcast_v20f16_to_v40i8(<20 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 @@ -22467,17 +22508,17 @@ define <40 x i8> @bitcast_v20f16_to_v40i8(<20 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v17, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v18, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB60_2: ; %Flow @@ -22501,17 +22542,17 @@ define <40 x i8> @bitcast_v20f16_to_v40i8(<20 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v17, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v18, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB60_4: ; %end @@ -22519,100 +22560,105 @@ define <40 x i8> @bitcast_v20f16_to_v40i8(<20 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v2.h, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v1.h, v11.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v29 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v16, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v28, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v14, v15 ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v4.l, v11.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v29, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v7.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v11.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v30, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v26, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v13, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v17, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v19, v10 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v15 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v17.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v11, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v15 ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[10:11], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v20f16_to_v40i8: @@ -28813,50 +28859,50 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:24 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:32 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:20 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:4 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.h, v29.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v27.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v25.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v48.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v38.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v39.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v38.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(8) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v36.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v36.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v36.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v36.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(6) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v38.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v49 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB72_3 @@ -28871,228 +28917,243 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v34.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v34.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v30.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v0.h, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v24.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v26.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v21.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v5.h, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v16.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v5.h, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v17 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v27.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v1.h, v27.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v2.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v22.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v3.l, v23.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v28.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v4.l, v21.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v21.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v5.l, v19.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v6.h, v19.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v6.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v7.l, v17.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v31.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v11, v10 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v8.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v9.l, v16.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v10 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB72_2 ; GFX11-TRUE16-NEXT: .LBB72_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v35.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v34.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v28.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v30.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v33.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v29.h, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v26.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v34.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v30.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v34.l, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v27.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v23.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v27.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v22.h, 3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v28.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v23.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v25.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v21.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v24.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v25.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v29.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v26.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v29.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v27.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v24.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v25.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v27.h, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v23.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v23.h, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v11 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v21.h, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v22.h, v4.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v21.l, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v19.h, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v18.h, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v19.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v21.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v19.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v20.h, v6.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v30.l, 3 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v31.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v19.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v18.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v11 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v16.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v16.h, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v17.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v17.h, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v18.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v17.l, v7.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v17 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v17.h, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v16.h, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v16.l, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v11 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -30847,20 +30908,20 @@ define <40 x i8> @bitcast_v5f64_to_v40i8(<5 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 @@ -30876,17 +30937,17 @@ define <40 x i8> @bitcast_v5f64_to_v40i8(<5 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v17, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v18, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB74_2: ; %Flow @@ -30905,17 +30966,17 @@ define <40 x i8> @bitcast_v5f64_to_v40i8(<5 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v17, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v18, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB74_4: ; %end @@ -30923,100 +30984,105 @@ define <40 x i8> @bitcast_v5f64_to_v40i8(<5 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v2.h, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v1.h, v11.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v29 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v16, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v28, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v14, v15 ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v4.l, v11.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v29, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v7.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v11.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v30, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v26, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v13, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v17, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v19, v10 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v15 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v17.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v11, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v15 ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[10:11], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v5f64_to_v40i8: @@ -32944,50 +33010,50 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:24 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:32 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:20 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:4 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.h, v29.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v27.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v25.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v48.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v38.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v39.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v38.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(8) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v36.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v36.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v36.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v36.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(6) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v38.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v49 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB76_3 @@ -33002,228 +33068,243 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v34.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v34.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v30.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v0.h, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v24.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v26.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v21.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v5.h, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v16.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v5.h, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v17 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v27.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v1.h, v27.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v2.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v22.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v3.l, v23.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v28.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v4.l, v21.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v21.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v5.l, v19.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v6.h, v19.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v6.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v7.l, v17.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v31.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v11, v10 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v8.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v9.l, v16.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v10 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB76_2 ; GFX11-TRUE16-NEXT: .LBB76_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v35.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v34.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v28.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v30.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v33.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v29.h, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v26.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v34.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v30.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v34.l, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v27.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v23.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v27.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v22.h, 3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v28.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v23.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v25.l, v1.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v21.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v24.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v25.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v29.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v26.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v29.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v27.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v24.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v25.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v27.h, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v23.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v23.h, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v11 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v21.h, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v22.h, v4.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v21.l, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v19.h, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v18.h, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v19.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v21.l, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v19.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v20.h, v6.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v30.l, 3 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v31.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v19.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v18.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v11 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v16.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v16.h, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v17.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v17.h, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v18.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v3, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v17.l, v7.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v14, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v17 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v17.h, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v16.h, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v16.l, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v11 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -34993,20 +35074,20 @@ define <40 x i8> @bitcast_v5i64_to_v40i8(<5 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 @@ -35022,17 +35103,17 @@ define <40 x i8> @bitcast_v5i64_to_v40i8(<5 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v17, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v18, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB78_2: ; %Flow @@ -35059,17 +35140,17 @@ define <40 x i8> @bitcast_v5i64_to_v40i8(<5 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[15:16], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v17, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v18, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v20, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v21, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v22, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v24, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1 ; GFX11-TRUE16-NEXT: .LBB78_4: ; %end @@ -35077,100 +35158,105 @@ define <40 x i8> @bitcast_v5i64_to_v40i8(<5 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v2.h, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v1.h, v11.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v29 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v16, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v28, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v14, v15 ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v4.l, v11.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v29, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v7.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v14, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v13, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v11.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v8.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v30, v28 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v26, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v16, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v13, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v17, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v19, v10 +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v13 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v15 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v17.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v11, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v11, v15 ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[10:11], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v5i64_to_v40i8: diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll index bfed1f4..48c9b87 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll @@ -2257,8 +2257,8 @@ define i32 @bitcast_v4i8_to_i32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4 @@ -2273,18 +2273,19 @@ define i32 @bitcast_v4i8_to_i32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB22_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB22_2 ; GFX11-TRUE16-NEXT: .LBB22_4: ; %cmp.true @@ -2294,16 +2295,17 @@ define i32 @bitcast_v4i8_to_i32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -4504,8 +4506,8 @@ define float @bitcast_v4i8_to_f32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4 @@ -4520,18 +4522,19 @@ define float @bitcast_v4i8_to_f32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB42_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB42_2 ; GFX11-TRUE16-NEXT: .LBB42_4: ; %cmp.true @@ -4541,16 +4544,17 @@ define float @bitcast_v4i8_to_f32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -6463,8 +6467,8 @@ define <2 x i16> @bitcast_v4i8_to_v2i16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4 @@ -6479,18 +6483,19 @@ define <2 x i16> @bitcast_v4i8_to_v2i16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB58_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB58_2 ; GFX11-TRUE16-NEXT: .LBB58_4: ; %cmp.true @@ -6500,16 +6505,17 @@ define <2 x i16> @bitcast_v4i8_to_v2i16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -8110,8 +8116,8 @@ define <2 x half> @bitcast_v4i8_to_v2f16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4 @@ -8126,18 +8132,19 @@ define <2 x half> @bitcast_v4i8_to_v2f16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2 ; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true @@ -8147,16 +8154,17 @@ define <2 x half> @bitcast_v4i8_to_v2f16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -9471,8 +9479,8 @@ define <2 x bfloat> @bitcast_v4i8_to_v2bf16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4 @@ -9487,18 +9495,19 @@ define <2 x bfloat> @bitcast_v4i8_to_v2bf16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB78_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB78_2 ; GFX11-TRUE16-NEXT: .LBB78_4: ; %cmp.true @@ -9508,16 +9517,17 @@ define <2 x bfloat> @bitcast_v4i8_to_v2bf16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -10183,8 +10193,8 @@ define <1 x i32> @bitcast_v4i8_to_v1i32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4 @@ -10199,18 +10209,19 @@ define <1 x i32> @bitcast_v4i8_to_v1i32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB82_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB82_2 ; GFX11-TRUE16-NEXT: .LBB82_4: ; %cmp.true @@ -10220,16 +10231,17 @@ define <1 x i32> @bitcast_v4i8_to_v1i32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll index 45e205b..68312b8 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.48bit.ll @@ -145,37 +145,36 @@ define <3 x half> @bitcast_v3bf16_to_v3f16(<3 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB0_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 ; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v1, 16, 1 ; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v0, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v0 ; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v0, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v3, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v1 ; GFX11-TRUE16-NEXT: v_add3_u32 v5, v6, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h -; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v4, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0x7fc0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v2, v0 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v3, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.h ; GFX11-TRUE16-NEXT: .LBB0_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 @@ -797,40 +796,40 @@ define <3 x i16> @bitcast_v3bf16_to_v3i16(<3 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB4_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff0000, v0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v0 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v3, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v1, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v5, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v4, v6, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 16, v1 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshl_or_b32 v0, v0, 16, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshl_or_b32 v1, 0x7fc0, 16, v1 +; GFX11-TRUE16-NEXT: v_lshl_or_b32 v0, v0, 16, v1 +; GFX11-TRUE16-NEXT: v_lshl_or_b32 v1, 0x7fc0, 16, v2 ; GFX11-TRUE16-NEXT: .LBB4_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll index d524514..5aac06a 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll @@ -8768,32 +8768,32 @@ define <64 x i8> @bitcast_v16i32_to_v64i8(<16 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 @@ -8812,26 +8812,26 @@ define <64 x i8> @bitcast_v16i32_to_v64i8(<16 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB24_2: ; %Flow @@ -8864,26 +8864,26 @@ define <64 x i8> @bitcast_v16i32_to_v64i8(<16 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB24_4: ; %end @@ -8891,151 +8891,156 @@ define <64 x i8> @bitcast_v16i32_to_v64i8(<16 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v54, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v22 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v53, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v8.h, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v21, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24 ; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v34 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v13.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v49, v50 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v51, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v35, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v22, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v19, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v23, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v25, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24 ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -12465,13 +12470,13 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:128 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:124 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:120 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:116 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:112 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:108 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:104 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:100 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:92 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:88 ; GFX11-TRUE16-NEXT: scratch_load_b32 v82, off, s32 offset:132 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 @@ -12487,81 +12492,83 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:80 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:84 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:60 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:52 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v25.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.h, v8.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v80.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v65.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(25) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v66.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(23) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v66.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(21) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v67.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(20) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v67.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(19) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v68.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v68.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v69.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(15) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v70.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(14) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v70.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(12) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v71.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v80.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v82 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB26_3 @@ -12574,366 +12581,384 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v51.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v53.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v53.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v54.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v50.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v51.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v52.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v2.h, v51.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v6, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v39.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v48.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v48.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v0.h, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v50.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v51.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v48.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v49.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v1.h, v50.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v2.l, v49.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v3.l, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v25.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v24.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v26.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v36.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v13, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v11.h, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v20.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v16.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v11.h, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v23, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v18 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v38.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v4.l, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v29.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v5.l, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v26.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v38.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v6.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v37.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v7.l, v23.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v37.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v35.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v8.l, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v35.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v9.l, v21.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v21.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v10.l, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v11.l, v19.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v12.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v13.l, v17.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v31.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v14.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v54 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v15.l, v16.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v54 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2 ; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v54.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v53.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v51.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v53.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v50.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v49.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v54.h, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v52.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v48.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v53.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v50.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v52.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v51.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v52.l, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v51.l, v0.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v50.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v29.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v28.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v27.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v30.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v39.l, v1.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v52.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v50.h, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v39.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v30.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v55 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v49.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v51.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v39.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v48.l, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v48.h, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v48.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v27.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v49.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v27.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v24.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v29.h, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v25.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v29.l, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v28.h, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v23.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v9 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v24.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v38.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v26.h, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v36.h, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v26.h, v6.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v38.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v25.h, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v21.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v22.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v36.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v21.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v22.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v24.l, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v23.h, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v13, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v22.h, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v36.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l ; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v23.l, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v21.h, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v55 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v20.h, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v21.l, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v34.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v11.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v18.h, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v19.h, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v20.h, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v20.l, v12.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v33.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v31.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v19.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v18.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v55 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v32.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v16.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v16.h, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v17.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v17.h, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v17.l, v13.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v23, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v18 +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.h, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v14.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v55 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v55 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -23563,32 +23588,32 @@ define <64 x i8> @bitcast_v16f32_to_v64i8(<16 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 @@ -23607,26 +23632,26 @@ define <64 x i8> @bitcast_v16f32_to_v64i8(<16 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB48_2: ; %Flow @@ -23651,26 +23676,26 @@ define <64 x i8> @bitcast_v16f32_to_v64i8(<16 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB48_4: ; %end @@ -23678,151 +23703,156 @@ define <64 x i8> @bitcast_v16f32_to_v64i8(<16 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v54, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v22 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v53, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v8.h, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v21, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24 ; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v34 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v13.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v49, v50 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v51, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v35, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v22, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v19, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v23, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v25, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24 ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -27383,13 +27413,13 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:128 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:124 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:120 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:116 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:112 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:108 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:104 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:100 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:92 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:88 ; GFX11-TRUE16-NEXT: scratch_load_b32 v82, off, s32 offset:132 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 @@ -27405,81 +27435,83 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:80 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:84 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:60 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:52 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v25.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.h, v8.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v80.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v65.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(25) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v66.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(23) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v66.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(21) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v67.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(20) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v67.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(19) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v68.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v68.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v69.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(15) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v70.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(14) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v70.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(12) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v71.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v80.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v82 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB50_3 @@ -27492,366 +27524,384 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v51.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v53.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v53.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v54.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v50.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v51.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v52.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v2.h, v51.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v6, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v39.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v48.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v48.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v0.h, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v50.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v51.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v48.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v49.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v1.h, v50.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v2.l, v49.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v3.l, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v25.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v24.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v26.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v36.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v13, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v11.h, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v20.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v16.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v11.h, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v23, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v18 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v38.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v4.l, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v29.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v5.l, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v26.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v38.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v6.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v37.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v7.l, v23.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v37.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v35.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v8.l, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v35.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v9.l, v21.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v21.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v10.l, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v11.l, v19.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v12.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v13.l, v17.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v31.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v14.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v54 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v15.l, v16.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v54 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2 ; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v54.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v53.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v51.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v53.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v50.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v49.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v54.h, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v52.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v48.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v53.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v50.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v52.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v51.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v52.l, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v51.l, v0.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v50.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v29.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v28.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v27.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v30.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v39.l, v1.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v52.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v50.h, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v39.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v30.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v55 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v49.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v51.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v39.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v48.l, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v48.h, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v48.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v27.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v49.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v27.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v24.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v29.h, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v25.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v29.l, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v28.h, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v23.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v9 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v24.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v38.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v26.h, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v36.h, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v26.h, v6.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v38.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v25.h, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v21.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v22.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v36.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v21.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v22.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v24.l, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v23.h, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v13, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v22.h, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v36.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l ; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v23.l, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v21.h, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v55 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v20.h, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v21.l, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v34.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v11.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v18.h, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v19.h, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v20.h, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v20.l, v12.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v33.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v31.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v19.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v18.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v55 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v32.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v16.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v16.h, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v17.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v17.h, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v17.l, v13.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v23, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v18 +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.h, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v14.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v55 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v55 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -37866,32 +37916,32 @@ define <64 x i8> @bitcast_v8i64_to_v64i8(<8 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 @@ -37910,26 +37960,26 @@ define <64 x i8> @bitcast_v8i64_to_v64i8(<8 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB68_2: ; %Flow @@ -37967,26 +38017,26 @@ define <64 x i8> @bitcast_v8i64_to_v64i8(<8 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB68_4: ; %end @@ -37994,151 +38044,156 @@ define <64 x i8> @bitcast_v8i64_to_v64i8(<8 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v54, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v22 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v53, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v8.h, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v21, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24 ; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v34 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v13.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v49, v50 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v51, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v35, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v22, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v19, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v23, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v25, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24 ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -41573,13 +41628,13 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:128 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:124 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:120 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:116 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:112 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:108 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:104 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:100 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:92 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:88 ; GFX11-TRUE16-NEXT: scratch_load_b32 v82, off, s32 offset:132 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 @@ -41595,81 +41650,83 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:80 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:84 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:60 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:52 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v25.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.h, v8.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v80.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v65.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(25) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v66.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(23) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v66.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(21) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v67.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(20) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v67.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(19) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v68.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v68.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v69.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(15) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v70.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(14) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v70.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(12) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v71.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v80.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v82 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB70_3 @@ -41682,366 +41739,384 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v51.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v53.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v53.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v54.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v50.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v51.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v52.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v2.h, v51.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v6, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v39.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v48.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v48.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v0.h, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v50.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v51.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v48.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v49.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v1.h, v50.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v2.l, v49.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v3.l, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v25.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v24.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v26.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v36.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v13, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v11.h, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v20.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v16.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v11.h, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v23, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v18 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v38.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v4.l, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v29.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v5.l, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v26.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v38.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v6.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v37.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v7.l, v23.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v37.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v35.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v8.l, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v35.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v9.l, v21.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v21.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v10.l, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v11.l, v19.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v12.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v13.l, v17.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v31.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v14.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v54 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v15.l, v16.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v54 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2 ; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v54.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v53.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v51.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v53.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v50.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v49.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v54.h, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v52.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v48.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v53.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v50.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v52.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v51.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v52.l, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v51.l, v0.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v50.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v29.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v28.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v27.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v30.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v39.l, v1.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v52.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v50.h, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v39.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v30.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v55 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v49.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v51.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v39.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v48.l, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v48.h, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v48.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v27.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v49.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v27.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v24.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v29.h, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v25.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v29.l, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v28.h, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v23.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v9 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v24.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v38.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v26.h, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v36.h, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v26.h, v6.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v38.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v25.h, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v21.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v22.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v36.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v21.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v22.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v24.l, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v23.h, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v13, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v22.h, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v36.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l ; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v23.l, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v21.h, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v55 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v20.h, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v21.l, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v34.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v11.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v18.h, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v19.h, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v20.h, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v20.l, v12.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v33.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v31.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v19.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v18.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v55 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v32.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v16.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v16.h, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v17.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v17.h, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v17.l, v13.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v23, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v18 +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.h, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v14.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v55 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v55 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -51220,32 +51295,32 @@ define <64 x i8> @bitcast_v8f64_to_v64i8(<8 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 @@ -51264,26 +51339,26 @@ define <64 x i8> @bitcast_v8f64_to_v64i8(<8 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB84_2: ; %Flow @@ -51308,26 +51383,26 @@ define <64 x i8> @bitcast_v8f64_to_v64i8(<8 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB84_4: ; %end @@ -51335,151 +51410,156 @@ define <64 x i8> @bitcast_v8f64_to_v64i8(<8 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v54, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v22 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v53, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v8.h, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v21, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24 ; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v34 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v13.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v49, v50 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v51, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v35, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v22, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v19, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v23, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v25, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24 ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -54909,13 +54989,13 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:128 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:124 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:120 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:116 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:112 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:108 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:104 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:100 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:92 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:88 ; GFX11-TRUE16-NEXT: scratch_load_b32 v82, off, s32 offset:132 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 @@ -54931,81 +55011,83 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:80 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:84 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:60 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:52 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v25.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.h, v8.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v80.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v65.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(25) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v66.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(23) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v66.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(21) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v67.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(20) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v67.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(19) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v68.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v68.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v69.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(15) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v70.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(14) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v70.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(12) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v71.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v80.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v82 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB86_3 @@ -55018,366 +55100,384 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v51.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v53.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v53.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v54.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v50.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v51.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v52.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v2.h, v51.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v6, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v39.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.h, v48.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v48.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v0.h, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v50.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v51.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v48.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v49.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v1.h, v50.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v2.l, v49.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v3.l, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v25.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v24.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v26.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v36.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v6.h, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.h, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v13, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v11.h, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v20.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v16.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v11.h, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v18.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v23, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v18 +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v38.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v4.l, v29.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v29.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v5.l, v28.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v26.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v38.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v6.l, v25.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v37.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v7.l, v23.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v37.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v35.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v8.l, v22.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v35.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v9.l, v21.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v21.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v10.l, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v11.l, v19.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v12.l, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v54 +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v13.l, v17.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v31.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v14.l, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v54 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v15.l, v16.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v54 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2 ; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v54.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v53.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v51.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v53.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v50.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v49.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v54.h, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v52.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v48.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v53.h, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v50.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v52.l, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v51.h, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v52.l, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v51.l, v0.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v50.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v29.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v28.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v27.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v30.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v39.l, v1.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v52.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v50.h, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v39.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v30.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v55 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v49.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v51.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v39.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v48.l, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v48.h, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v48.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v27.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v49.l, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v27.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v24.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v29.h, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v25.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v29.l, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v28.h, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v23.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v9 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v24.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v5, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v38.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v26.h, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v36.h, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v26.h, v6.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v38.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v25.h, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v21.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v22.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v36.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v21.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v22.h, v7.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v24.l, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v23.h, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v11, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v13, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v22.h, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v36.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l ; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v23.l, v8.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v21.h, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v55 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v20.h, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v21.l, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v34.h, 3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.h, 3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v55 +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v11.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v18.h, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v19.h, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v20.h, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v20.l, v12.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v33.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v32.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v31.l, 3 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v19.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v18.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v55 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v32.h, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v16.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v16.h, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v17.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v17.h, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v15 +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v17.l, v13.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v13 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v23, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v14, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v15, v18 +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.h, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v14.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v17 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v55 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v55 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -64473,32 +64573,32 @@ define <64 x i8> @bitcast_v32i16_to_v64i8(<32 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 @@ -64517,26 +64617,26 @@ define <64 x i8> @bitcast_v32i16_to_v64i8(<32 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB96_2: ; %Flow @@ -64569,26 +64669,26 @@ define <64 x i8> @bitcast_v32i16_to_v64i8(<32 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB96_4: ; %end @@ -64596,151 +64696,156 @@ define <64 x i8> @bitcast_v32i16_to_v64i8(<32 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v54, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v22 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v53, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v8.h, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v21, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24 ; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v34 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v13.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v49, v50 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v51, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v35, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v22, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v19, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v23, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v25, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24 ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -76596,32 +76701,32 @@ define <64 x i8> @bitcast_v32f16_to_v64i8(<32 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 @@ -76640,26 +76745,26 @@ define <64 x i8> @bitcast_v32f16_to_v64i8(<32 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB104_2: ; %Flow @@ -76692,26 +76797,26 @@ define <64 x i8> @bitcast_v32f16_to_v64i8(<32 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1 ; GFX11-TRUE16-NEXT: .LBB104_4: ; %end @@ -76719,151 +76824,156 @@ define <64 x i8> @bitcast_v32f16_to_v64i8(<32 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25 ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v54, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v22 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v53, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v8.h, v20.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v34, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v21, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24 ; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v34 -; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v13.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21 +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v49, v50 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v51, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v35, v11 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v21, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v22, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v19, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v23, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v25, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24 ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -85582,58 +85692,58 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) @@ -85648,297 +85758,304 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 24, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 24, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v67, 8, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v1 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v2.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.h, v3.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.h, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.h, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, v6.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v8.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v8.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v10.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v10.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.h, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.h, v13.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.h, v11.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.h, v13.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.h, v14.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v84.h, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.h, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.h, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.h, v14.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.h, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.h, v16.h ; GFX11-TRUE16-NEXT: .LBB108_2: ; %Flow ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB108_4 ; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true -; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v4 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_lshlrev_b32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v11 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v1, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v17, 0x40c00000, v17 -; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v17, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v17 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 -; GFX11-TRUE16-NEXT: v_add3_u32 v21, v21, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v2, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v2 ; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v17, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v21, v21, v2, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v27, v20, v22, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_add_f32 v1, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v27.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v26, v20, v22, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18 -; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v26.h -; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v27, v21, v23, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v26, v21, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18 +; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v1, 0x7fff ; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v27 +; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff +; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v26 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v28, v20, v21, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 ; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v4, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v2 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v28.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v17, v22, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v19, 16, 1 ; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v4, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v4 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v19, 0x7fff -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v20, v1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v19, 16, 1 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v2 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v30, v18, v22, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 ; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v21 :: v_dual_lshlrev_b32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v29, v17, v23, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v30.h +; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v19, 0x7fff +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 ; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v21, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v21 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v29, v17, v23 ; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v3, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v21, 0x7fff -; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v21 ; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v31, v18, v19 :: v_dual_add_f32 v18, 0x40c00000, v22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v30.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v32, v18, v19, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v31.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v18, 16, 1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v4, v23, vcc_lo +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v20, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v32.h +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v4, v23 :: v_dual_add_f32 v18, 0x40c00000, v22 ; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v17, v29 ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v6, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v19, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v18, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v18, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18 ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v6, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v6 -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v20, 16, 1 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v32, v17, v21, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v18, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v33, v17, v21, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v20, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v33, v19, v22, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v31, v19, v22 :: v_dual_and_b32 v20, 0xffff0000, v5 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v20, 0x40c00000, v20 :: v_dual_lshlrev_b32 v5, 16, v5 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v8, 16, v8 +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v20, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v20 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v32.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_add_f32 v5, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v33.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v5, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v20, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v35, v17, v21, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v17, v21, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v8, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v35.h +; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v36.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v6, v22, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v19, 16, 1 ; GFX11-TRUE16-NEXT: v_add3_u32 v6, v17, v8, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v8 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v34, v6, v17 :: v_dual_add_f32 v19, 0x40c00000, v19 -; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v18, v33 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v34.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v19, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v19 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 24, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v6 ; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v19, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v36, v5, v22 :: v_dual_and_b32 v23, 0xffff0000, v7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v35, v6, v17, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v7 +; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v18, v31 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v34, v5, v22, vcc_lo ; GFX11-TRUE16-NEXT: v_bfi_b32 v5, 0xffff, v21, v20 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v36 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v35.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_dual_add_f32 v20, 0x40c00000, v20 :: v_dual_add_f32 v7, 0x40c00000, v7 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v8 +; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v34 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v7, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v7 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v8 ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v7, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v10 ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v20, 16, 1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v49, v19, v21, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v49, v19, v21, vcc_lo ; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v23 :: v_dual_add_f32 v10, 0x40c00000, v10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v20, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v9 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v17, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v17 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v10, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v17, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v10, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v18, v22, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v10 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v49.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v38, v7, v21, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v39, v7, v21, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v12 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v12 ; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v48, v19, v22 :: v_dual_lshlrev_b32 v7, 16, v9 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v38.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_add_f32 v7, 0x40c00000, v7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v10 -; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v19, v48 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v12, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v7, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h +; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfe_u32 v22, v21, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v21 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v7 -; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v7, 0x7fff +; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v19, v48 +; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v12, 16, 1 ; GFX11-TRUE16-NEXT: v_add3_u32 v22, v22, v21, 0x7fff -; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v23 -; GFX11-TRUE16-NEXT: v_add3_u32 v24, v24, v12, 0x7fff +; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v50, 0x400000, v12 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v39 :: v_dual_cndmask_b32 v52, v22, v37 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v10 +; GFX11-TRUE16-NEXT: v_add3_u32 v24, v24, v12, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v54, v22, v37, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v7, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v7 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v9, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v9 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v14 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v52.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 +; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v7, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v54.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_lshlrev_b32 v11, 16, v11 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v65, v19, v25, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 -; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v9, 0x7fff -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v9 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v38 +; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v14, 16, 1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v52, v24, v50 :: v_dual_add_f32 v9, 0x40c00000, v23 ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v21, 16, 1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v53, v24, v50, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v7, v52 +; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v9, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v9 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v14, 0x40c00000, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v10 -; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v7, v53 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v20, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v11 ; GFX11-TRUE16-NEXT: v_add3_u32 v11, v19, v21, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v9, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v21 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v67, 8, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v20, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21 -; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v14, 16, 1 ; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v7, 16, 1 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v65.h -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 24, v12 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v11, v11, v19, vcc_lo ; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v21, v23, v7, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v7 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v13 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 ; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v19, 16, 1 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v20, v9 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v23 :: v_dual_cndmask_b32 v70, v21, v22 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v23 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v71, v21, v22, vcc_lo ; GFX11-TRUE16-NEXT: v_add3_u32 v21, v24, v19, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v19 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 ; GFX11-TRUE16-NEXT: v_add3_u32 v23, v25, v14, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v14 ; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v7, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v67, v21, v22, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_add_f32 v13, 0x40c00000, v13 :: v_dual_cndmask_b32 v66, v21, v22 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v14, v25, v7, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v16 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v68, v23, v24, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 ; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v13, 16, 1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 8, v12 +; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v20, v9 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v16, 0x40c00000, v16 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v68.h ; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v7, v14, v19 :: v_dual_add_f32 v14, 0x40c00000, v21 @@ -85950,42 +86067,42 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v15 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v16 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v85, v19, v23, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v86, v19, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v16, 0x7fff ; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v21, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 ; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v14, 16, 1 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v15 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v21 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v21 ; GFX11-TRUE16-NEXT: v_add3_u32 v23, v23, v21, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v81, v13, v25, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v82, v13, v25, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21 ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v24, v14, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v14 ; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v15, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v84, v23, v39, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v85, v23, v38, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v85.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v86.h ; GFX11-TRUE16-NEXT: v_add3_u32 v13, v37, v15, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v70.h -; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v22, v67 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v82, v19, v24, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v71.h +; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v22, v66 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v81, v19, v24, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v81.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v84.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v82.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v85.h ; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v23, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 24, v14 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v13, v13, v16, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v16, 0xffff, v19, v82 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v14 +; GFX11-TRUE16-NEXT: v_bfi_b32 v16, 0xffff, v19, v81 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v9 ; GFX11-TRUE16-NEXT: v_bfi_b32 v15, 0xffff, v15, v13 ; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v21, v7 ; GFX11-TRUE16-NEXT: v_bfi_b32 v7, 0xffff, v18, v17 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12] @@ -85994,160 +86111,159 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v7 ; GFX11-TRUE16-NEXT: .LBB108_4: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v28.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v113.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v112.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v26.h ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v1.h, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v102.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v103.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v24.l ; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v2.h, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v101.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.h, v5.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.h, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v14, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v49.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v102.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v101.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v24 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.h, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v99.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v100.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v98.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v97.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v49.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v31.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v96.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v12 ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v34.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v87.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v16, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.h, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v80.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v71.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v35.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v87.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v65.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v83.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v80.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v10 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v10.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v24 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v71.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v48.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v14, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v54.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v11.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v23, v24 -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v70.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v69.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v16, v21 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v22 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v14, v19 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v64.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v86.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v14, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v12.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v52.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v53.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v16, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v53.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v16, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v19 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v14.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v85.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v55.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v68.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.h, v15.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v16.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.h, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v23 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v13 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v67.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v84.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v81.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v82.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v51.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v50.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v66.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v82.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v14.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19 ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v25, v6 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v20 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v21, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v22, v17 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v23, v24 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v25, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24 ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll index a40ee16..6fe6665 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll @@ -2160,46 +2160,47 @@ define i64 @bitcast_v4bf16_to_i64(<4 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB22_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v1, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v1, v7, v10 :: v_dual_lshlrev_b32 v0, 16, v0 +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v9, v11 :: v_dual_and_b32 v0, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v2, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v0, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v0, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v2, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v8, v10, vcc_lo +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v3, v4 ; GFX11-TRUE16-NEXT: .LBB22_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -3064,12 +3065,13 @@ define i64 @bitcast_v8i8_to_i64(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8 @@ -3083,66 +3085,61 @@ define i64 @bitcast_v8i8_to_i64(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2 ; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -5307,46 +5304,47 @@ define double @bitcast_v4bf16_to_f64(<4 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB46_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v1, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v1, v7, v10 :: v_dual_lshlrev_b32 v0, 16, v0 +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v9, v11 :: v_dual_and_b32 v0, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v2, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v0, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v0, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v2, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v8, v10, vcc_lo +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v3, v4 ; GFX11-TRUE16-NEXT: .LBB46_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -6216,12 +6214,13 @@ define double @bitcast_v8i8_to_f64(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8 @@ -6235,66 +6234,61 @@ define double @bitcast_v8i8_to_f64(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2 ; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -8166,46 +8160,47 @@ define <2 x i32> @bitcast_v4bf16_to_v2i32(<4 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB66_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v1, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v1, v7, v10 :: v_dual_lshlrev_b32 v0, 16, v0 +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v9, v11 :: v_dual_and_b32 v0, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v2, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v0, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v0, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v2, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v8, v10, vcc_lo +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v3, v4 ; GFX11-TRUE16-NEXT: .LBB66_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -9068,12 +9063,13 @@ define <2 x i32> @bitcast_v8i8_to_v2i32(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8 @@ -9087,66 +9083,61 @@ define <2 x i32> @bitcast_v8i8_to_v2i32(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2 ; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -10698,46 +10689,47 @@ define <2 x float> @bitcast_v4bf16_to_v2f32(<4 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB82_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v1, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v1, v7, v10 :: v_dual_lshlrev_b32 v0, 16, v0 +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v9, v11 :: v_dual_and_b32 v0, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v2, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v0, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v0, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v2, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v8, v10, vcc_lo +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v3, v4 ; GFX11-TRUE16-NEXT: .LBB82_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -11611,12 +11603,13 @@ define <2 x float> @bitcast_v8i8_to_v2f32(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8 @@ -11630,66 +11623,61 @@ define <2 x float> @bitcast_v8i8_to_v2f32(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2 ; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -12867,49 +12855,49 @@ define <4 x i16> @bitcast_v4bf16_to_v4i16(<4 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB94_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v4, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v4 ; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v1, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v4, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v1, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_add_f32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v2, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 +; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v2, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v0, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v0 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v0, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v6, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 16, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v10, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 16, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v3, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.h ; GFX11-TRUE16-NEXT: v_lshl_or_b32 v1, v0, 16, v1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshl_or_b32 v0, v3, 16, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_lshl_or_b32 v0, v2, 16, v3 ; GFX11-TRUE16-NEXT: .LBB94_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -13841,12 +13829,13 @@ define <4 x i16> @bitcast_v8i8_to_v4i16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8 @@ -13860,66 +13849,61 @@ define <4 x i16> @bitcast_v8i8_to_v4i16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB98_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_2 ; GFX11-TRUE16-NEXT: .LBB98_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -14696,46 +14680,46 @@ define <4 x half> @bitcast_v4bf16_to_v4f16(<4 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB102_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v1, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v1, v6, v7 :: v_dual_add_f32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v3, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v9, v11 :: v_dual_and_b32 v0, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v9, v11 :: v_dual_add_f32 v2, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v2, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v0, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h ; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v3, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v8, v10, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v3, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v2, v0 ; GFX11-TRUE16-NEXT: .LBB102_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -15671,12 +15655,13 @@ define <4 x half> @bitcast_v8i8_to_v4f16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8 @@ -15690,66 +15675,61 @@ define <4 x half> @bitcast_v8i8_to_v4f16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB106_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB106_2 ; GFX11-TRUE16-NEXT: .LBB106_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -16347,42 +16327,41 @@ define <8 x i8> @bitcast_v4bf16_to_v8i8(<4 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB108_4 ; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v8.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v9 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v0 :: v_dual_lshlrev_b32 v0, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_add_f32 v1, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v5, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v12, v4, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v7, v10, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v0 :: v_dual_add_f32 v0, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v9.l +; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v0, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_add3_u32 v11, v4, v3, 0x7fff ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v9, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v3, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v6, v0, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v2, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v4, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v1, 0x7fff ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v12, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v2, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v9, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v11, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v8, v10, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v3, v2 -; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v1, v6 +; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v2, v1 +; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v3, v6 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[10:11], 24, v[8:9] @@ -16987,12 +16966,13 @@ define <4 x bfloat> @bitcast_v8i8_to_v4bf16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8 @@ -17006,66 +16986,61 @@ define <4 x bfloat> @bitcast_v8i8_to_v4bf16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB110_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB110_2 ; GFX11-TRUE16-NEXT: .LBB110_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll index 5163539..e5245f7 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll @@ -1102,15 +1102,15 @@ define <3 x i32> @bitcast_v12i8_to_v3i32(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v0.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v11.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12 @@ -1126,76 +1126,79 @@ define <3 x i32> @bitcast_v12i8_to_v3i32(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v0.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v3.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v2 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v2 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v5 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v2.l, v3.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v7 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB6_2 ; GFX11-TRUE16-NEXT: .LBB6_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v7.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v7.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v6.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v5.h, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.l, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v4.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v1.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v7 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -2084,62 +2087,57 @@ define <3 x i32> @bitcast_v6bf16_to_v3i32(<6 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB10_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v3, 16, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_add_f32 v2, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v0, v1, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_cndmask_b32 v5, v7, v9 +; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v11, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v6, v8 :: v_dual_and_b32 v2, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_add3_u32 v8, v10, v4, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v0, 16, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v9, vcc_lo -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v3, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v6, v10, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v7, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v11 +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v2, 16, 1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v6 :: v_dual_cndmask_b32 v0, v0, v8 +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v11, 0x7fff +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v10, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v7, v7, v9, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v10, v5, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v11, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v1, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v10, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v5 +; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v3, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v6, v1 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc_lo +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v4, v5 ; GFX11-TRUE16-NEXT: .LBB10_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -4243,15 +4241,15 @@ define <3 x float> @bitcast_v12i8_to_v3f32(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v0.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v11.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12 @@ -4267,76 +4265,79 @@ define <3 x float> @bitcast_v12i8_to_v3f32(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v0.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v3.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v2 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v2 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v5 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v2.l, v3.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v7 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB22_2 ; GFX11-TRUE16-NEXT: .LBB22_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v7.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v7.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v6.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v5.h, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.l, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v4.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v1.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v7 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v7 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -5229,62 +5230,57 @@ define <3 x float> @bitcast_v6bf16_to_v3f32(<6 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v3, 16, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_add_f32 v2, 0x40c00000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v0, v1, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_cndmask_b32 v5, v7, v9 +; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v11, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v6, v8 :: v_dual_and_b32 v2, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_add3_u32 v8, v10, v4, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_lshlrev_b32 v0, 16, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v9, vcc_lo -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v3, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v6, v10, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v7, v12, vcc_lo -; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v11 +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v2, 16, 1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v6 :: v_dual_cndmask_b32 v0, v0, v8 +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v11, 0x7fff +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v10, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v7, v7, v9, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v10, v5, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v11, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v1, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v10, vcc_lo -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v5 +; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v3, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v6, v1 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc_lo +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v4, v5 ; GFX11-TRUE16-NEXT: .LBB26_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -6889,16 +6885,16 @@ define <6 x bfloat> @bitcast_v12i8_to_v6bf16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v11.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12 @@ -6914,77 +6910,79 @@ define <6 x bfloat> @bitcast_v12i8_to_v6bf16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v8.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v6.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v2 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v2 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v5 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v2.l, v4.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v7 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB36_2 ; GFX11-TRUE16-NEXT: .LBB36_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v9.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v8.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v6.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.h, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v6.l, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v1.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.h, v2.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v9 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -7783,67 +7781,68 @@ define <12 x i8> @bitcast_v6bf16_to_v12i8(<6 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB38_4 ; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v13 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v10 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_add_f32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v12 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v13.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v3, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v12.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v2, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v10.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v10 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 +; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v11, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v0, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v11, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v10, v4, v0, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff0000, v13 -; GFX11-TRUE16-NEXT: v_bfe_u32 v0, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add3_u32 v0, v0, v3, 0x7fff ; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v11, v4, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v11 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v10, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v0, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.h +; GFX11-TRUE16-NEXT: v_add3_u32 v13, v0, v1, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v4.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v11, v9, vcc_lo -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v0, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v12 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v7, 16, 1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v8, v13, vcc_lo -; GFX11-TRUE16-NEXT: v_add3_u32 v3, v12, v7, 0x7fff -; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v1, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v2 +; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v13, v14, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v10, v9, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v5, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v3, v12, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.h +; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v7, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v7, 24, v13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v2, v1 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v3, v11, vcc_lo ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v11, 0x7fc07fc0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v8.h -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, 8, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v8.h ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[14:15], 24, v[12:13] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v12 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v7, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v7, 24, v13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v5, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, 8, v13 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[16:17], 24, v[10:11] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v9, 8, v10 ; GFX11-TRUE16-NEXT: .LBB38_4: ; %end @@ -8652,16 +8651,16 @@ define <6 x half> @bitcast_v12i8_to_v6f16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v11.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12 @@ -8677,77 +8676,79 @@ define <6 x half> @bitcast_v12i8_to_v6f16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v8.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v6.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v2 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v2 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v5 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v2.l, v4.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v7 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB40_2 ; GFX11-TRUE16-NEXT: .LBB40_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v9.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v8.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v6.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.h, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v6.l, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v1.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.h, v2.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v9 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -10064,16 +10065,16 @@ define <6 x i16> @bitcast_v12i8_to_v6i16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v11.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12 @@ -10089,77 +10090,79 @@ define <6 x i16> @bitcast_v12i8_to_v6i16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v8.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v6.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v2 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v2 +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v4.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v5 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v2.l, v4.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v7 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB44_2 ; GFX11-TRUE16-NEXT: .LBB44_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v9.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v8.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v6.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.h, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v6.l, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v1.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.h, v2.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v9 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -11446,59 +11449,61 @@ define <6 x half> @bitcast_v6bf16_to_v6f16(<6 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB48_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_add_f32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_add_f32 v6, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v1.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v6, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v1, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v4, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_cndmask_b32 v1, v8, v2 +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v0, 16, 1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v9, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v1, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v7, v9, vcc_lo -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v10, v4, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v6, v10, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v3 -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v2, v8, v12 :: v_dual_add_f32 v5, 0x40c00000, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v4 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v10, v0, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v9, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v10, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v7, v12 :: v_dual_add_f32 v5, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v5, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v7, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v11, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v6, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v11, v5, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v10, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc_lo +; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v6, v0 +; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v7, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v4, v5 ; GFX11-TRUE16-NEXT: .LBB48_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] @@ -12390,64 +12395,66 @@ define <6 x i16> @bitcast_v6bf16_to_v6i16(<6 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB52_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_add_f32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_lshlrev_b32 v4, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v2, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v9, v4, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v4, v6, v8 :: v_dual_add_f32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: v_add3_u32 v8, v11, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: v_add3_u32 v11, v12, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v9, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v4, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v9 +; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v9, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v5, v8, v9 :: v_dual_and_b32 v0, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v11, v12, vcc_lo +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v6 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v10, v9, 0x7fff +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v11, v3, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_add3_u32 v7, v10, v1, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_add3_u32 v11, v12, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v9, v10 :: v_dual_add_f32 v0, 0x40c00000, v0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v0, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v11, v12, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v6, v13, vcc_lo +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v0, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v7, v13, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v7, v10, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 16, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v8, v2, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 16, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, 16, v6 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.h ; GFX11-TRUE16-NEXT: v_lshl_or_b32 v2, v0, 16, v2 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshl_or_b32 v1, v1, 16, v4 -; GFX11-TRUE16-NEXT: v_lshl_or_b32 v0, v5, 16, v3 +; GFX11-TRUE16-NEXT: v_lshl_or_b32 v1, v1, 16, v3 +; GFX11-TRUE16-NEXT: v_lshl_or_b32 v0, v5, 16, v4 ; GFX11-TRUE16-NEXT: .LBB52_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] diff --git a/llvm/test/CodeGen/AMDGPU/atomicrmw-bf16-gfx11plus.ll b/llvm/test/CodeGen/AMDGPU/atomicrmw-bf16-gfx11plus.ll new file mode 100644 index 0000000..535f05bc --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/atomicrmw-bf16-gfx11plus.ll @@ -0,0 +1,122 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4 +; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck --check-prefixes=GFX11-TRUE16 %s +; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck --check-prefixes=GFX11-FAKE16 %s + +@global_smem = external local_unnamed_addr addrspace(1) global [0 x i8], align 16 + +define amdgpu_kernel void @v_atomicrmw_fadd_bf16(ptr addrspace(1) %out, i1 %in, ptr addrspace(1) %ptr) #0 { +; GFX11-TRUE16-LABEL: v_atomicrmw_fadd_bf16: +; GFX11-TRUE16: ; %bb.0: +; GFX11-TRUE16-NEXT: s_clause 0x1 +; GFX11-TRUE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x34 +; GFX11-TRUE16-NEXT: s_load_b64 s[2:3], s[4:5], 0x24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0x3ff, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, 0 +; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-TRUE16-NEXT: global_load_b32 v2, v0, s[0:1] offset:4 +; GFX11-TRUE16-NEXT: s_and_b32 s0, s2, -4 +; GFX11-TRUE16-NEXT: s_mov_b32 s1, s3 +; GFX11-TRUE16-NEXT: s_and_b32 s2, s2, 3 +; GFX11-TRUE16-NEXT: s_load_b32 s3, s[0:1], 0x0 +; GFX11-TRUE16-NEXT: s_lshl_b32 s2, s2, 3 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h +; GFX11-TRUE16-NEXT: s_lshl_b32 s4, 0xffff, s2 +; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, s3 +; GFX11-TRUE16-NEXT: s_not_b32 s3, s4 +; GFX11-TRUE16-NEXT: s_mov_b32 s4, 0 +; GFX11-TRUE16-NEXT: .p2align 6 +; GFX11-TRUE16-NEXT: .LBB0_1: ; %atomicrmw.start +; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s2, v1 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, v0, v2 +; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v0, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s2, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_or_b32 v0, v1, s3, v0 +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v0, v4, v[0:1], s[0:1] glc +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: buffer_gl1_inv +; GFX11-TRUE16-NEXT: buffer_gl0_inv +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, v0 +; GFX11-TRUE16-NEXT: s_or_b32 s4, vcc_lo, s4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4 +; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB0_1 +; GFX11-TRUE16-NEXT: ; %bb.2: ; %atomicrmw.end +; GFX11-TRUE16-NEXT: s_endpgm +; +; GFX11-FAKE16-LABEL: v_atomicrmw_fadd_bf16: +; GFX11-FAKE16: ; %bb.0: +; GFX11-FAKE16-NEXT: s_clause 0x1 +; GFX11-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x34 +; GFX11-FAKE16-NEXT: s_load_b64 s[2:3], s[4:5], 0x24 +; GFX11-FAKE16-NEXT: v_dual_mov_b32 v3, 0 :: v_dual_and_b32 v0, 0x3ff, v0 +; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-FAKE16-NEXT: global_load_b32 v0, v0, s[0:1] offset:4 +; GFX11-FAKE16-NEXT: s_and_b32 s0, s2, -4 +; GFX11-FAKE16-NEXT: s_mov_b32 s1, s3 +; GFX11-FAKE16-NEXT: s_and_b32 s2, s2, 3 +; GFX11-FAKE16-NEXT: s_load_b32 s3, s[0:1], 0x0 +; GFX11-FAKE16-NEXT: s_lshl_b32 s2, s2, 3 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; GFX11-FAKE16-NEXT: s_lshl_b32 s4, 0xffff, s2 +; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-FAKE16-NEXT: v_mov_b32_e32 v1, s3 +; GFX11-FAKE16-NEXT: s_not_b32 s3, s4 +; GFX11-FAKE16-NEXT: s_mov_b32 s4, 0 +; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v0, 16, v0 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v2, 16, v0 +; GFX11-FAKE16-NEXT: .p2align 6 +; GFX11-FAKE16-NEXT: .LBB0_1: ; %atomicrmw.start +; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v0, s2, v1 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_add_f32_e32 v0, v0, v2 +; GFX11-FAKE16-NEXT: v_bfe_u32 v4, v0, 16, 1 +; GFX11-FAKE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 +; GFX11-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_add3_u32 v4, v4, v0, 0x7fff +; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v0, 16, v0 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v0, s2, v0 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_and_or_b32 v0, v1, s3, v0 +; GFX11-FAKE16-NEXT: global_atomic_cmpswap_b32 v0, v3, v[0:1], s[0:1] glc +; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-FAKE16-NEXT: buffer_gl1_inv +; GFX11-FAKE16-NEXT: buffer_gl0_inv +; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v0, v1 +; GFX11-FAKE16-NEXT: v_mov_b32_e32 v1, v0 +; GFX11-FAKE16-NEXT: s_or_b32 s4, vcc_lo, s4 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; GFX11-FAKE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s4 +; GFX11-FAKE16-NEXT: s_cbranch_execnz .LBB0_1 +; GFX11-FAKE16-NEXT: ; %bb.2: ; %atomicrmw.end +; GFX11-FAKE16-NEXT: s_endpgm + %tid = call i32 @llvm.amdgcn.workitem.id.x() + %in.gep = getelementptr <{ [0 x i8] }>, ptr addrspace(1) %ptr, i64 0, i32 0, i32 %tid + %load = load <4 x bfloat>, ptr addrspace(1) %in.gep + %extract1 = extractelement <4 x bfloat> %load, i64 3 + %fadd = atomicrmw fadd ptr addrspace(1) %out, bfloat %extract1 syncscope("agent") acq_rel + ret void +} diff --git a/llvm/test/CodeGen/AMDGPU/bf16.ll b/llvm/test/CodeGen/AMDGPU/bf16.ll index 2bdf994..cd6d741 100644 --- a/llvm/test/CodeGen/AMDGPU/bf16.ll +++ b/llvm/test/CodeGen/AMDGPU/bf16.ll @@ -9082,17 +9082,19 @@ define bfloat @v_fadd_bf16(bfloat %a, bfloat %b) { ; GFX11TRUE16-LABEL: v_fadd_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_add_f32_e32 v0, v0, v1 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_add_f32_e32 v0, v1, v2 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -13318,9 +13320,10 @@ define bfloat @v_fadd_bf16_fpimm_0(bfloat %arg0) { ; GFX11TRUE16-LABEL: v_fadd_bf16_fpimm_0: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_add_f32_e32 v0, 1.0, v0 +; GFX11TRUE16-NEXT: v_add_f32_e32 v0, 1.0, v1 ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 @@ -13413,9 +13416,10 @@ define bfloat @v_fadd_bf16_fpimm_1(bfloat %arg0) { ; GFX11TRUE16-LABEL: v_fadd_bf16_fpimm_1: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_add_f32_e32 v0, 0x42280000, v0 +; GFX11TRUE16-NEXT: v_add_f32_e32 v0, 0x42280000, v1 ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 @@ -13515,17 +13519,19 @@ define bfloat @v_fsub_bf16(bfloat %a, bfloat %b) { ; GFX11TRUE16-LABEL: v_fsub_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_sub_f32_e32 v0, v1, v2 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -14275,17 +14281,19 @@ define bfloat @v_fmul_bf16(bfloat %a, bfloat %b) { ; GFX11TRUE16-LABEL: v_fmul_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_mul_f32_e32 v0, v1, v2 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -18568,32 +18576,34 @@ define bfloat @v_fdiv_bf16(bfloat %a, bfloat %b) { ; GFX11TRUE16-LABEL: v_fdiv_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_div_scale_f32 v2, null, v1, v1, v0 -; GFX11TRUE16-NEXT: v_rcp_f32_e32 v3, v2 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.l +; GFX11TRUE16-NEXT: v_div_scale_f32 v1, null, v0, v0, v2 +; GFX11TRUE16-NEXT: v_div_scale_f32 v5, vcc_lo, v2, v0, v2 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_rcp_f32_e32 v3, v1 ; GFX11TRUE16-NEXT: s_waitcnt_depctr 0xfff -; GFX11TRUE16-NEXT: v_fma_f32 v4, -v2, v3, 1.0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_fma_f32 v4, -v1, v3, 1.0 ; GFX11TRUE16-NEXT: v_fmac_f32_e32 v3, v4, v3 -; GFX11TRUE16-NEXT: v_div_scale_f32 v5, vcc_lo, v0, v1, v0 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_mul_f32_e32 v4, v5, v3 +; GFX11TRUE16-NEXT: v_fma_f32 v6, -v1, v4, v5 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_fma_f32 v6, -v2, v4, v5 ; GFX11TRUE16-NEXT: v_fmac_f32_e32 v4, v6, v3 +; GFX11TRUE16-NEXT: v_fma_f32 v1, -v1, v4, v5 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_fma_f32 v2, -v2, v4, v5 -; GFX11TRUE16-NEXT: v_div_fmas_f32 v2, v2, v3, v4 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_div_fixup_f32 v0, v2, v1, v0 +; GFX11TRUE16-NEXT: v_div_fmas_f32 v1, v1, v3, v4 +; GFX11TRUE16-NEXT: v_div_fixup_f32 v0, v1, v0, v2 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -19018,17 +19028,19 @@ define bfloat @v_minnum_bf16(bfloat %a, bfloat %b) { ; GFX11TRUE16-LABEL: v_minnum_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_min_f32_e32 v0, v0, v1 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_min_f32_e32 v0, v1, v2 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -23270,17 +23282,19 @@ define bfloat @v_maxnum_bf16(bfloat %a, bfloat %b) { ; GFX11TRUE16-LABEL: v_maxnum_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_max_f32_e32 v0, v0, v1 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_max_f32_e32 v0, v1, v2 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -27591,11 +27605,12 @@ define bfloat @v_sqrt_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_sqrt_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11TRUE16-NEXT: v_mul_f32_e32 v1, 0x4f800000, v0 -; GFX11TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xf800000, v0 -; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xf800000, v1 +; GFX11TRUE16-NEXT: v_mul_f32_e32 v0, 0x4f800000, v1 +; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc_lo ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) ; GFX11TRUE16-NEXT: v_sqrt_f32_e32 v1, v0 ; GFX11TRUE16-NEXT: s_waitcnt_depctr 0xfff @@ -27730,9 +27745,10 @@ define bfloat @v_ldexp_bf16_i32(bfloat %a, i32 %b) { ; GFX11TRUE16-LABEL: v_ldexp_bf16_i32: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v0, v1 +; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v2, v1 ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 @@ -27836,17 +27852,18 @@ define { bfloat, i16 } @v_frexp_bf16_i16(bfloat %a) { ; GFX11TRUE16-LABEL: v_frexp_bf16_i16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) ; GFX11TRUE16-NEXT: v_frexp_mant_f32_e32 v0, v1 +; GFX11TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v1, v1 ; GFX11TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff ; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo -; GFX11TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v1, v1 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -28019,11 +28036,12 @@ define bfloat @v_log_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_log_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 -; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 32, vcc_lo -; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v0, v1 +; GFX11TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v1 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 32, vcc_lo +; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v1, v0 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_log_f32_e32 v0, v0 ; GFX11TRUE16-NEXT: s_waitcnt_depctr 0xfff @@ -28177,13 +28195,14 @@ define bfloat @v_log2_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_log2_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX11TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 -; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 32, vcc_lo +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v1 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 32, vcc_lo +; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v1, v0 ; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x42000000, vcc_lo -; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v0, v2 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_log_f32_e32 v0, v0 ; GFX11TRUE16-NEXT: s_waitcnt_depctr 0xfff ; GFX11TRUE16-NEXT: v_sub_f32_e32 v0, v0, v1 @@ -28367,11 +28386,12 @@ define bfloat @v_log10_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_log10_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 -; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 32, vcc_lo -; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v0, v1 +; GFX11TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v1 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 32, vcc_lo +; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v1, v0 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_log_f32_e32 v0, v0 ; GFX11TRUE16-NEXT: s_waitcnt_depctr 0xfff @@ -28580,25 +28600,26 @@ define bfloat @v_exp_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_exp_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_mul_f32_e32 v1, 0x3fb8aa3b, v0 -; GFX11TRUE16-NEXT: v_rndne_f32_e32 v2, v1 -; GFX11TRUE16-NEXT: v_fma_f32 v3, 0x3fb8aa3b, v0, -v1 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11TRUE16-NEXT: v_mul_f32_e32 v0, 0x3fb8aa3b, v1 +; GFX11TRUE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, 0xc2ce8ed0, v1 +; GFX11TRUE16-NEXT: v_fma_f32 v2, 0x3fb8aa3b, v1, -v0 +; GFX11TRUE16-NEXT: v_rndne_f32_e32 v3, v0 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11TRUE16-NEXT: v_sub_f32_e32 v1, v1, v2 -; GFX11TRUE16-NEXT: v_fmamk_f32 v3, v0, 0x32a5705f, v3 -; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v2, v2 -; GFX11TRUE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, 0xc2ce8ed0, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_add_f32_e32 v1, v1, v3 -; GFX11TRUE16-NEXT: v_exp_f32_e32 v1, v1 +; GFX11TRUE16-NEXT: v_fmamk_f32 v2, v1, 0x32a5705f, v2 +; GFX11TRUE16-NEXT: v_sub_f32_e32 v0, v0, v3 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11TRUE16-NEXT: v_add_f32_e32 v0, v0, v2 +; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v2, v3 +; GFX11TRUE16-NEXT: v_exp_f32_e32 v0, v0 ; GFX11TRUE16-NEXT: s_waitcnt_depctr 0xfff -; GFX11TRUE16-NEXT: v_ldexp_f32 v1, v1, v2 +; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v0, v2 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc_lo -; GFX11TRUE16-NEXT: v_cmp_nlt_f32_e32 vcc_lo, 0x42b17218, v0 -; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, 0x7f800000, v1, vcc_lo +; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc_lo +; GFX11TRUE16-NEXT: v_cmp_nlt_f32_e32 vcc_lo, 0x42b17218, v1 +; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, 0x7f800000, v0, vcc_lo ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 @@ -28744,13 +28765,14 @@ define bfloat @v_exp2_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_exp2_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX11TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v0 -; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0x42800000, vcc_lo +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v1 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 0x42800000, vcc_lo +; GFX11TRUE16-NEXT: v_add_f32_e32 v0, v1, v0 ; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0xffffffc0, vcc_lo -; GFX11TRUE16-NEXT: v_add_f32_e32 v0, v0, v2 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_exp_f32_e32 v0, v0 ; GFX11TRUE16-NEXT: s_waitcnt_depctr 0xfff ; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v0, v1 @@ -28937,25 +28959,26 @@ define bfloat @v_exp10_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_exp10_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_mul_f32_e32 v1, 0x40549a78, v0 -; GFX11TRUE16-NEXT: v_rndne_f32_e32 v2, v1 -; GFX11TRUE16-NEXT: v_fma_f32 v3, 0x40549a78, v0, -v1 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11TRUE16-NEXT: v_mul_f32_e32 v0, 0x40549a78, v1 +; GFX11TRUE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, 0xc23369f4, v1 +; GFX11TRUE16-NEXT: v_fma_f32 v2, 0x40549a78, v1, -v0 +; GFX11TRUE16-NEXT: v_rndne_f32_e32 v3, v0 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11TRUE16-NEXT: v_sub_f32_e32 v1, v1, v2 -; GFX11TRUE16-NEXT: v_fmamk_f32 v3, v0, 0x33979a37, v3 -; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v2, v2 -; GFX11TRUE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, 0xc23369f4, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_add_f32_e32 v1, v1, v3 -; GFX11TRUE16-NEXT: v_exp_f32_e32 v1, v1 +; GFX11TRUE16-NEXT: v_fmamk_f32 v2, v1, 0x33979a37, v2 +; GFX11TRUE16-NEXT: v_sub_f32_e32 v0, v0, v3 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11TRUE16-NEXT: v_add_f32_e32 v0, v0, v2 +; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v2, v3 +; GFX11TRUE16-NEXT: v_exp_f32_e32 v0, v0 ; GFX11TRUE16-NEXT: s_waitcnt_depctr 0xfff -; GFX11TRUE16-NEXT: v_ldexp_f32 v1, v1, v2 +; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v0, v2 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc_lo -; GFX11TRUE16-NEXT: v_cmp_nlt_f32_e32 vcc_lo, 0x421a209b, v0 -; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, 0x7f800000, v1, vcc_lo +; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, 0, v0, vcc_lo +; GFX11TRUE16-NEXT: v_cmp_nlt_f32_e32 vcc_lo, 0x421a209b, v1 +; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, 0x7f800000, v0, vcc_lo ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 @@ -29066,9 +29089,10 @@ define bfloat @v_ceil_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_ceil_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_ceil_f32_e32 v0, v0 +; GFX11TRUE16-NEXT: v_ceil_f32_e32 v0, v1 ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 @@ -29163,9 +29187,10 @@ define bfloat @v_trunc_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_trunc_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_trunc_f32_e32 v0, v0 +; GFX11TRUE16-NEXT: v_trunc_f32_e32 v0, v1 ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 @@ -29260,9 +29285,10 @@ define bfloat @v_rint_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_rint_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_rndne_f32_e32 v0, v0 +; GFX11TRUE16-NEXT: v_rndne_f32_e32 v0, v1 ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 @@ -29357,9 +29383,10 @@ define bfloat @v_nearbyint_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_nearbyint_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_rndne_f32_e32 v0, v0 +; GFX11TRUE16-NEXT: v_rndne_f32_e32 v0, v1 ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 @@ -29483,16 +29510,17 @@ define bfloat @v_round_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_round_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_trunc_f32_e32 v1, v0 -; GFX11TRUE16-NEXT: v_sub_f32_e32 v2, v0, v1 +; GFX11TRUE16-NEXT: v_trunc_f32_e32 v0, v1 +; GFX11TRUE16-NEXT: v_sub_f32_e32 v2, v1, v0 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_cmp_ge_f32_e64 s0, |v2|, 0.5 ; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 1.0, s0 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_bfi_b32 v0, 0x7fffffff, v2, v0 -; GFX11TRUE16-NEXT: v_add_f32_e32 v0, v1, v0 +; GFX11TRUE16-NEXT: v_bfi_b32 v1, 0x7fffffff, v2, v1 +; GFX11TRUE16-NEXT: v_add_f32_e32 v0, v0, v1 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 @@ -29594,9 +29622,10 @@ define bfloat @v_roundeven_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_roundeven_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_rndne_f32_e32 v0, v0 +; GFX11TRUE16-NEXT: v_rndne_f32_e32 v0, v1 ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 @@ -29691,9 +29720,10 @@ define bfloat @v_floor_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_floor_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_floor_f32_e32 v0, v0 +; GFX11TRUE16-NEXT: v_floor_f32_e32 v0, v1 ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 @@ -29786,9 +29816,10 @@ define bfloat @v_canonicalize_bf16(bfloat %a) { ; GFX11TRUE16-LABEL: v_canonicalize_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_max_f32_e32 v0, v0, v0 +; GFX11TRUE16-NEXT: v_max_f32_e32 v0, v1, v1 ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 @@ -29916,15 +29947,27 @@ define i1 @v_fcmp_oeq_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_oeq_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_eq_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_oeq_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_oeq_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp oeq bfloat %a, %b ret i1 %op } @@ -29979,15 +30022,27 @@ define i1 @v_fcmp_ogt_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_ogt_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_gt_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_ogt_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_ogt_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp ogt bfloat %a, %b ret i1 %op } @@ -30042,15 +30097,27 @@ define i1 @v_fcmp_oge_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_oge_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_ge_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_oge_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_ge_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_oge_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_ge_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp oge bfloat %a, %b ret i1 %op } @@ -30105,15 +30172,27 @@ define i1 @v_fcmp_olt_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_olt_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_lt_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_olt_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_olt_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp olt bfloat %a, %b ret i1 %op } @@ -30168,15 +30247,27 @@ define i1 @v_fcmp_ole_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_ole_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_le_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_ole_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_le_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_ole_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_le_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp ole bfloat %a, %b ret i1 %op } @@ -30231,15 +30322,27 @@ define i1 @v_fcmp_one_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_one_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_lg_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_one_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_lg_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_one_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_lg_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp one bfloat %a, %b ret i1 %op } @@ -30294,15 +30397,27 @@ define i1 @v_fcmp_uno_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_uno_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_uno_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_uno_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp uno bfloat %a, %b ret i1 %op } @@ -30357,15 +30472,27 @@ define i1 @v_fcmp_ueq_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_ueq_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_nlg_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_ueq_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_nlg_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_ueq_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_nlg_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp ueq bfloat %a, %b ret i1 %op } @@ -30420,15 +30547,27 @@ define i1 @v_fcmp_ugt_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_ugt_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_nle_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_ugt_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_nle_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_ugt_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_nle_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp ugt bfloat %a, %b ret i1 %op } @@ -30483,15 +30622,27 @@ define i1 @v_fcmp_uge_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_uge_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_nlt_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_uge_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_nlt_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_uge_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_nlt_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp uge bfloat %a, %b ret i1 %op } @@ -30546,15 +30697,27 @@ define i1 @v_fcmp_ult_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_ult_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_nge_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_ult_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_nge_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_ult_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_nge_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp ult bfloat %a, %b ret i1 %op } @@ -30609,15 +30772,27 @@ define i1 @v_fcmp_ule_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_ule_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_ule_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_ule_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp ule bfloat %a, %b ret i1 %op } @@ -30672,15 +30847,27 @@ define i1 @v_fcmp_une_bf16(bfloat %a, bfloat %b) { ; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fcmp_une_bf16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cmp_neq_f32_e32 vcc_lo, v0, v1 -; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fcmp_une_bf16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11TRUE16-NEXT: v_cmp_neq_f32_e32 vcc_lo, v1, v2 +; GFX11TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fcmp_une_bf16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cmp_neq_f32_e32 vcc_lo, v0, v1 +; GFX11FAKE16-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fcmp une bfloat %a, %b ret i1 %op } @@ -30763,13 +30950,22 @@ define i16 @v_fptosi_bf16_to_i16(bfloat %x) { ; GFX10-NEXT: v_cvt_i32_f32_e32 v0, v0 ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fptosi_bf16_to_i16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cvt_i32_f32_e32 v0, v0 -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fptosi_bf16_to_i16: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v0, v1 +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fptosi_bf16_to_i16: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cvt_i32_f32_e32 v0, v0 +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fptosi bfloat %x to i16 ret i16 %op } @@ -31144,13 +31340,22 @@ define i32 @v_fptosi_bf16_to_i32(bfloat %x) { ; GFX10-NEXT: v_cvt_i32_f32_e32 v0, v0 ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fptosi_bf16_to_i32: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_cvt_i32_f32_e32 v0, v0 -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fptosi_bf16_to_i32: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v0, v1 +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fptosi_bf16_to_i32: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_cvt_i32_f32_e32 v0, v0 +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fptosi bfloat %x to i32 ret i32 %op } @@ -31494,27 +31699,50 @@ define i64 @v_fptosi_bf16_to_i64(bfloat %x) { ; GFX10-NEXT: v_sub_co_ci_u32_e32 v1, vcc_lo, v1, v3, vcc_lo ; GFX10-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: v_fptosi_bf16_to_i64: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-NEXT: v_trunc_f32_e32 v0, v0 -; GFX11-NEXT: v_mul_f32_e64 v1, 0x2f800000, |v0| -; GFX11-NEXT: v_ashrrev_i32_e32 v3, 31, v0 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-NEXT: v_floor_f32_e32 v1, v1 -; GFX11-NEXT: v_fma_f32 v2, 0xcf800000, v1, |v0| -; GFX11-NEXT: v_cvt_u32_f32_e32 v1, v1 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-NEXT: v_cvt_u32_f32_e32 v0, v2 -; GFX11-NEXT: v_xor_b32_e32 v1, v1, v3 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-NEXT: v_xor_b32_e32 v0, v0, v3 -; GFX11-NEXT: v_sub_co_u32 v0, vcc_lo, v0, v3 -; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NEXT: v_sub_co_ci_u32_e64 v1, null, v1, v3, vcc_lo -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11TRUE16-LABEL: v_fptosi_bf16_to_i64: +; GFX11TRUE16: ; %bb.0: +; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_trunc_f32_e32 v0, v1 +; GFX11TRUE16-NEXT: v_mul_f32_e64 v1, 0x2f800000, |v0| +; GFX11TRUE16-NEXT: v_ashrrev_i32_e32 v3, 31, v0 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_floor_f32_e32 v1, v1 +; GFX11TRUE16-NEXT: v_fma_f32 v2, 0xcf800000, v1, |v0| +; GFX11TRUE16-NEXT: v_cvt_u32_f32_e32 v1, v1 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11TRUE16-NEXT: v_cvt_u32_f32_e32 v0, v2 +; GFX11TRUE16-NEXT: v_xor_b32_e32 v1, v1, v3 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_xor_b32_e32 v0, v0, v3 +; GFX11TRUE16-NEXT: v_sub_co_u32 v0, vcc_lo, v0, v3 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_sub_co_ci_u32_e64 v1, null, v1, v3, vcc_lo +; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11FAKE16-LABEL: v_fptosi_bf16_to_i64: +; GFX11FAKE16: ; %bb.0: +; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_trunc_f32_e32 v0, v0 +; GFX11FAKE16-NEXT: v_mul_f32_e64 v1, 0x2f800000, |v0| +; GFX11FAKE16-NEXT: v_ashrrev_i32_e32 v3, 31, v0 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_floor_f32_e32 v1, v1 +; GFX11FAKE16-NEXT: v_fma_f32 v2, 0xcf800000, v1, |v0| +; GFX11FAKE16-NEXT: v_cvt_u32_f32_e32 v1, v1 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11FAKE16-NEXT: v_cvt_u32_f32_e32 v0, v2 +; GFX11FAKE16-NEXT: v_xor_b32_e32 v1, v1, v3 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_xor_b32_e32 v0, v0, v3 +; GFX11FAKE16-NEXT: v_sub_co_u32 v0, vcc_lo, v0, v3 +; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11FAKE16-NEXT: v_sub_co_ci_u32_e64 v1, null, v1, v3, vcc_lo +; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31] %op = fptosi bfloat %x to i64 ret i64 %op } @@ -42575,18 +42803,21 @@ define bfloat @v_fma_bf16(bfloat %a, bfloat %b, bfloat %c) { ; GFX11TRUE16-LABEL: v_fma_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.l +; GFX11TRUE16-NEXT: v_fmac_f32_e32 v3, v1, v2 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11TRUE16-NEXT: v_bfe_u32 v0, v3, 16, 1 +; GFX11TRUE16-NEXT: v_or_b32_e32 v1, 0x400000, v3 +; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11TRUE16-NEXT: v_add3_u32 v0, v0, v3, 0x7fff ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_fmac_f32_e32 v2, v0, v1 -; GFX11TRUE16-NEXT: v_bfe_u32 v0, v2, 16, 1 -; GFX11TRUE16-NEXT: v_or_b32_e32 v1, 0x400000, v2 -; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_add3_u32 v0, v0, v2, 0x7fff ; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -43457,26 +43688,30 @@ define bfloat @v_fmuladd_bf16(bfloat %a, bfloat %b, bfloat %c) { ; GFX11TRUE16-LABEL: v_fmuladd_bf16: ; GFX11TRUE16: ; %bb.0: ; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.l +; GFX11TRUE16-NEXT: v_mul_f32_e32 v0, v1, v3 +; GFX11TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 -; GFX11TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v0 +; GFX11TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff -; GFX11TRUE16-NEXT: v_dual_cndmask_b32 v0, v1, v3 :: v_dual_lshlrev_b32 v1, 16, v2 ; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc_lo ; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11TRUE16-NEXT: v_add_f32_e32 v0, v0, v1 -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11TRUE16-NEXT: v_add_f32_e32 v0, v0, v3 ; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff -; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo +; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll index 348862d..f4b432d 100644 --- a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll +++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll @@ -5100,55 +5100,56 @@ define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__amdgpu_no_fine ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: s_addk_co_i32 s16, 0x200 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, -4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, s4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v5, s4 ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX12-TRUE16-NEXT: buffer_load_b32 v1, v4, s[0:3], null offen +; GFX12-TRUE16-NEXT: buffer_load_b32 v2, v5, s[0:3], null offen ; GFX12-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX12-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX12-TRUE16-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v0, v0, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX12-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v1, v1, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX12-TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0 -; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], null offen th:TH_ATOMIC_RETURN +; GFX12-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1 +; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], null offen th:TH_ATOMIC_RETURN ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v3 ; GFX12-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB16_1 ; GFX12-TRUE16-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX12-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s5 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v2 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v3 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -5257,53 +5258,54 @@ define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__amdgpu_no_fine ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: s_addk_i32 s16, 0x200 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, -4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, s4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v5, s4 ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX11-TRUE16-NEXT: buffer_load_b32 v1, v4, s[0:3], 0 offen +; GFX11-TRUE16-NEXT: buffer_load_b32 v2, v5, s[0:3], 0 offen ; GFX11-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX11-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, v1, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, v0, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v2 -; GFX11-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v3 +; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0 -; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], 0 offen glc +; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1 +; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], 0 offen glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB16_1 ; GFX11-TRUE16-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v3 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__amdgpu_no_fine_grained_memory: @@ -5619,48 +5621,49 @@ define void @buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset__amdgpu_no_fine ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: s_addk_co_i32 s16, 0x200 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, -4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, s4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, s4 ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX12-TRUE16-NEXT: buffer_load_b32 v1, v2, s[0:3], null offen +; GFX12-TRUE16-NEXT: buffer_load_b32 v2, v3, s[0:3], null offen ; GFX12-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX12-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX12-TRUE16-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v0, v0, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v4, v0, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX12-TRUE16-NEXT: v_add3_u32 v4, v4, v0, 0x7fff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v1, v1, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v4.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v1, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v1, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v4 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0 -; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], null offen th:TH_ATOMIC_RETURN +; GFX12-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1 +; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v1, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v4 ; GFX12-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 @@ -5773,46 +5776,47 @@ define void @buffer_fat_ptr_agent_atomic_fadd_noret_bf16__offset__amdgpu_no_fine ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: s_addk_i32 s16, 0x200 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, -4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, s4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, s4 ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX11-TRUE16-NEXT: buffer_load_b32 v1, v2, s[0:3], 0 offen +; GFX11-TRUE16-NEXT: buffer_load_b32 v2, v3, s[0:3], 0 offen ; GFX11-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX11-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, v1, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v4.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v1, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, v0, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v4 -; GFX11-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v4 +; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0 -; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], 0 offen glc +; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1 +; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v4 ; GFX11-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 @@ -6124,15 +6128,15 @@ define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_nc_u32_e32 v4, 0x200, v4 +; GFX12-TRUE16-NEXT: v_add_nc_u32_e32 v6, 0x200, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s1, exec_lo ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 3, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, -4, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 3, v6 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v6 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, -4, v6 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v6, v7, 0xffff -; GFX12-TRUE16-NEXT: v_not_b32_e32 v9, v6 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v7, v4, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v11, v7 ; GFX12-TRUE16-NEXT: .LBB18_1: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 ; GFX12-TRUE16-NEXT: v_readfirstlane_b32 s5, v1 @@ -6146,39 +6150,38 @@ define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: buffer_load_b32 v6, v8, s[4:7], null offen +; GFX12-TRUE16-NEXT: buffer_load_b32 v7, v10, s[4:7], null offen ; GFX12-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB18_1 ; GFX12-TRUE16-NEXT: ; %bb.2: ; GFX12-TRUE16-NEXT: s_mov_b32 exec_lo, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 ; GFX12-TRUE16-NEXT: s_mov_b32 s1, 0 ; GFX12-TRUE16-NEXT: .LBB18_3: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: ; Child Loop BB18_4 Depth 2 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v7, v6 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v6, v4, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.h, v5.l ; GFX12-TRUE16-NEXT: s_mov_b32 s2, exec_lo ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v4, v4, v10 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff +; GFX12-TRUE16-NEXT: v_add_f32_e32 v6, v6, v8 +; GFX12-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v6 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v4, v5, v11, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v12, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v8.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, v4, v6 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v7, v5 -; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v9, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v5 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v5, v6 +; GFX12-TRUE16-NEXT: v_and_or_b32 v6, v7, v11, v6 +; GFX12-TRUE16-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6 ; GFX12-TRUE16-NEXT: .LBB18_4: ; Parent Loop BB18_3 Depth=1 ; GFX12-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2 ; GFX12-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 @@ -6193,14 +6196,14 @@ define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v8, s[4:7], null offen th:TH_ATOMIC_RETURN +; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[8:9], v10, s[4:7], null offen th:TH_ATOMIC_RETURN ; GFX12-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB18_4 ; GFX12-TRUE16-NEXT: ; %bb.5: ; in Loop: Header=BB18_3 Depth=1 ; GFX12-TRUE16-NEXT: s_mov_b32 exec_lo, s2 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v8, v7 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v7, v8 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_or_b32 s1, vcc_lo, s1 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -6208,7 +6211,7 @@ define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB18_3 ; GFX12-TRUE16-NEXT: ; %bb.6: ; %atomicrmw.end ; GFX12-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s1 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v7, v4 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v4, v8 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -6384,16 +6387,16 @@ define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 0x200, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 0x200, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s1, 0 ; GFX11-TRUE16-NEXT: s_mov_b32 s2, exec_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 3, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, -4, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 3, v6 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v6 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, -4, v6 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v6, v7, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v9, v6 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v7, v4, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v11, v7 ; GFX11-TRUE16-NEXT: .LBB18_1: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v1 @@ -6405,39 +6408,38 @@ define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-NEXT: s_and_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 -; GFX11-TRUE16-NEXT: buffer_load_b32 v6, v8, s[4:7], 0 offen +; GFX11-TRUE16-NEXT: buffer_load_b32 v7, v10, s[4:7], 0 offen ; GFX11-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB18_1 ; GFX11-TRUE16-NEXT: ; %bb.2: ; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 ; GFX11-TRUE16-NEXT: s_set_inst_prefetch_distance 0x1 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB18_3: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: ; Child Loop BB18_4 Depth 2 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v7, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v6, v4, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v5.l ; GFX11-TRUE16-NEXT: s_mov_b32 s2, exec_lo ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, v4, v10 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v5, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h +; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, v6, v8 +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v12, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, v4, v6 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v7, v5 -; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v9, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v5 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v5, v6 +; GFX11-TRUE16-NEXT: v_and_or_b32 v6, v7, v11, v6 +; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6 ; GFX11-TRUE16-NEXT: .LBB18_4: ; Parent Loop BB18_3 Depth=1 ; GFX11-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2 ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 @@ -6451,14 +6453,14 @@ define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v8, s[4:7], 0 offen glc +; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[8:9], v10, s[4:7], 0 offen glc ; GFX11-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB18_4 ; GFX11-TRUE16-NEXT: ; %bb.5: ; in Loop: Header=BB18_3 Depth=1 ; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s2 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v8, v7 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v7, v8 ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv ; GFX11-TRUE16-NEXT: s_or_b32 s1, vcc_lo, s1 @@ -6468,7 +6470,7 @@ define bfloat @buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-NEXT: ; %bb.6: ; %atomicrmw.end ; GFX11-TRUE16-NEXT: s_set_inst_prefetch_distance 0x2 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v7, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v4, v8 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: buffer_fat_ptr_agent_atomic_fadd_ret_bf16__offset__waterfall__amdgpu_no_fine_grained_memory: diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll index ab867b0..6f1675e 100644 --- a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll +++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll @@ -4228,55 +4228,56 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__amdgpu_no_fine ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: s_addk_co_i32 s16, 0x200 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, -4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, s4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v5, s4 ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX12-TRUE16-NEXT: buffer_load_b32 v1, v4, s[0:3], null offen +; GFX12-TRUE16-NEXT: buffer_load_b32 v2, v5, s[0:3], null offen ; GFX12-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX12-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX12-TRUE16-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v0, v0, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX12-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v1, v1, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX12-TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0 -; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], null offen th:TH_ATOMIC_RETURN +; GFX12-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1 +; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], null offen th:TH_ATOMIC_RETURN ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v3 ; GFX12-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB13_1 ; GFX12-TRUE16-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX12-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s5 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v2 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v3 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -4385,53 +4386,54 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__amdgpu_no_fine ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: s_addk_i32 s16, 0x200 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, -4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, s4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v5, s4 ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX11-TRUE16-NEXT: buffer_load_b32 v1, v4, s[0:3], 0 offen +; GFX11-TRUE16-NEXT: buffer_load_b32 v2, v5, s[0:3], 0 offen ; GFX11-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX11-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v1, v1, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v0, v0, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v2 -; GFX11-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v3 +; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0 -; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], 0 offen glc +; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1 +; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], 0 offen glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB13_1 ; GFX11-TRUE16-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v3 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__amdgpu_no_fine_grained_memory: @@ -4749,48 +4751,49 @@ define void @buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset__amdgpu_no_fine ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: s_addk_co_i32 s16, 0x200 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, -4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, s4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, s4 ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX12-TRUE16-NEXT: buffer_load_b32 v1, v2, s[0:3], null offen +; GFX12-TRUE16-NEXT: buffer_load_b32 v2, v3, s[0:3], null offen ; GFX12-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX12-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX12-TRUE16-NEXT: .LBB14_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v0, v0, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v4, v0, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX12-TRUE16-NEXT: v_add3_u32 v4, v4, v0, 0x7fff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v1, v1, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v4.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v1, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v1, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v4 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0 -; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], null offen th:TH_ATOMIC_RETURN +; GFX12-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1 +; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v1, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v4 ; GFX12-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 @@ -4903,46 +4906,47 @@ define void @buffer_fat_ptr_agent_atomic_fmax_noret_bf16__offset__amdgpu_no_fine ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: s_addk_i32 s16, 0x200 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, -4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, s4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, s4 ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX11-TRUE16-NEXT: buffer_load_b32 v1, v2, s[0:3], 0 offen +; GFX11-TRUE16-NEXT: buffer_load_b32 v2, v3, s[0:3], 0 offen ; GFX11-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX11-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB14_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v1, v1, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v4.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v1, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v0, v0, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v4 -; GFX11-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v4 +; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0 -; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], 0 offen glc +; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1 +; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v4 ; GFX11-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 @@ -5256,15 +5260,15 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_nc_u32_e32 v4, 0x200, v4 +; GFX12-TRUE16-NEXT: v_add_nc_u32_e32 v6, 0x200, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s1, exec_lo ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 3, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, -4, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 3, v6 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v6 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, -4, v6 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v6, v7, 0xffff -; GFX12-TRUE16-NEXT: v_not_b32_e32 v9, v6 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v7, v4, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v11, v7 ; GFX12-TRUE16-NEXT: .LBB15_1: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 ; GFX12-TRUE16-NEXT: v_readfirstlane_b32 s5, v1 @@ -5278,39 +5282,38 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: buffer_load_b32 v6, v8, s[4:7], null offen +; GFX12-TRUE16-NEXT: buffer_load_b32 v7, v10, s[4:7], null offen ; GFX12-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB15_1 ; GFX12-TRUE16-NEXT: ; %bb.2: ; GFX12-TRUE16-NEXT: s_mov_b32 exec_lo, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 ; GFX12-TRUE16-NEXT: s_mov_b32 s1, 0 ; GFX12-TRUE16-NEXT: .LBB15_3: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: ; Child Loop BB15_4 Depth 2 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v7, v6 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v6, v4, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.h, v5.l ; GFX12-TRUE16-NEXT: s_mov_b32 s2, exec_lo ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v4, v4, v10 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v6, v6, v8 +; GFX12-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v6 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v4, v5, v11, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v12, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v8.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, v4, v6 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v7, v5 -; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v9, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v5 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v5, v6 +; GFX12-TRUE16-NEXT: v_and_or_b32 v6, v7, v11, v6 +; GFX12-TRUE16-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6 ; GFX12-TRUE16-NEXT: .LBB15_4: ; Parent Loop BB15_3 Depth=1 ; GFX12-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2 ; GFX12-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 @@ -5325,14 +5328,14 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v8, s[4:7], null offen th:TH_ATOMIC_RETURN +; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[8:9], v10, s[4:7], null offen th:TH_ATOMIC_RETURN ; GFX12-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB15_4 ; GFX12-TRUE16-NEXT: ; %bb.5: ; in Loop: Header=BB15_3 Depth=1 ; GFX12-TRUE16-NEXT: s_mov_b32 exec_lo, s2 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v8, v7 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v7, v8 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_or_b32 s1, vcc_lo, s1 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -5340,7 +5343,7 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB15_3 ; GFX12-TRUE16-NEXT: ; %bb.6: ; %atomicrmw.end ; GFX12-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s1 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v7, v4 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v4, v8 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -5516,16 +5519,16 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 0x200, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 0x200, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s1, 0 ; GFX11-TRUE16-NEXT: s_mov_b32 s2, exec_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 3, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, -4, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 3, v6 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v6 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, -4, v6 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v6, v7, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v9, v6 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v7, v4, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v11, v7 ; GFX11-TRUE16-NEXT: .LBB15_1: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v1 @@ -5537,39 +5540,38 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-NEXT: s_and_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 -; GFX11-TRUE16-NEXT: buffer_load_b32 v6, v8, s[4:7], 0 offen +; GFX11-TRUE16-NEXT: buffer_load_b32 v7, v10, s[4:7], 0 offen ; GFX11-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB15_1 ; GFX11-TRUE16-NEXT: ; %bb.2: ; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 ; GFX11-TRUE16-NEXT: s_set_inst_prefetch_distance 0x1 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB15_3: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: ; Child Loop BB15_4 Depth 2 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v7, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v6, v4, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v5.l ; GFX11-TRUE16-NEXT: s_mov_b32 s2, exec_lo ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v4, v4, v10 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v5, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h +; GFX11-TRUE16-NEXT: v_max_f32_e32 v6, v6, v8 +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v12, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, v4, v6 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v7, v5 -; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v9, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v5 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v5, v6 +; GFX11-TRUE16-NEXT: v_and_or_b32 v6, v7, v11, v6 +; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6 ; GFX11-TRUE16-NEXT: .LBB15_4: ; Parent Loop BB15_3 Depth=1 ; GFX11-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2 ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 @@ -5583,14 +5585,14 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v8, s[4:7], 0 offen glc +; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[8:9], v10, s[4:7], 0 offen glc ; GFX11-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB15_4 ; GFX11-TRUE16-NEXT: ; %bb.5: ; in Loop: Header=BB15_3 Depth=1 ; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s2 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v8, v7 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v7, v8 ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv ; GFX11-TRUE16-NEXT: s_or_b32 s1, vcc_lo, s1 @@ -5600,7 +5602,7 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-NEXT: ; %bb.6: ; %atomicrmw.end ; GFX11-TRUE16-NEXT: s_set_inst_prefetch_distance 0x2 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v7, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v4, v8 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: buffer_fat_ptr_agent_atomic_fmax_ret_bf16__offset__waterfall__amdgpu_no_fine_grained_memory: diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll index 1a25904..acb27be 100644 --- a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll +++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll @@ -4228,55 +4228,56 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__amdgpu_no_fine ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: s_addk_co_i32 s16, 0x200 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, -4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, s4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v5, s4 ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX12-TRUE16-NEXT: buffer_load_b32 v1, v4, s[0:3], null offen +; GFX12-TRUE16-NEXT: buffer_load_b32 v2, v5, s[0:3], null offen ; GFX12-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX12-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX12-TRUE16-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v0, v0, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX12-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v1, v1, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX12-TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0 -; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], null offen th:TH_ATOMIC_RETURN +; GFX12-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1 +; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], null offen th:TH_ATOMIC_RETURN ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v3 ; GFX12-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB13_1 ; GFX12-TRUE16-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX12-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s5 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v2 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v3 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -4385,53 +4386,54 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__amdgpu_no_fine ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: s_addk_i32 s16, 0x200 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, -4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, s4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v5, s4 ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX11-TRUE16-NEXT: buffer_load_b32 v1, v4, s[0:3], 0 offen +; GFX11-TRUE16-NEXT: buffer_load_b32 v2, v5, s[0:3], 0 offen ; GFX11-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX11-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v1, v1, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v0, v0, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v2 -; GFX11-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v3 +; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0 -; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[2:3], v4, s[0:3], 0 offen glc +; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1 +; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], 0 offen glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB13_1 ; GFX11-TRUE16-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v3 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__amdgpu_no_fine_grained_memory: @@ -4749,48 +4751,49 @@ define void @buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset__amdgpu_no_fine ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: s_addk_co_i32 s16, 0x200 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, -4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, s4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, s4 ; GFX12-TRUE16-NEXT: s_and_b32 s4, s16, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX12-TRUE16-NEXT: buffer_load_b32 v1, v2, s[0:3], null offen +; GFX12-TRUE16-NEXT: buffer_load_b32 v2, v3, s[0:3], null offen ; GFX12-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX12-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX12-TRUE16-NEXT: .LBB14_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v0, v0, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v4, v0, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX12-TRUE16-NEXT: v_add3_u32 v4, v4, v0, 0x7fff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v1, v1, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v4.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v1, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v1, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v4 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe -; GFX12-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0 -; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], null offen th:TH_ATOMIC_RETURN +; GFX12-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1 +; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v1, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v4 ; GFX12-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 @@ -4903,46 +4906,47 @@ define void @buffer_fat_ptr_agent_atomic_fmin_noret_bf16__offset__amdgpu_no_fine ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: s_addk_i32 s16, 0x200 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, -4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, s4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, s4 ; GFX11-TRUE16-NEXT: s_and_b32 s4, s16, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s4, s4, 3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_lshl_b32 s5, 0xffff, s4 -; GFX11-TRUE16-NEXT: buffer_load_b32 v1, v2, s[0:3], 0 offen +; GFX11-TRUE16-NEXT: buffer_load_b32 v2, v3, s[0:3], 0 offen ; GFX11-TRUE16-NEXT: s_not_b32 s6, s5 ; GFX11-TRUE16-NEXT: s_mov_b32 s5, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB14_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s4, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v1, v1, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v4.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v1, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v0, v0, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v0, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s4, v4 -; GFX11-TRUE16-NEXT: v_and_or_b32 v0, v1, s6, v0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v4 +; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0 -; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v2, s[0:3], 0 offen glc +; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1 +; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v1 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v4 ; GFX11-TRUE16-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s5 @@ -5256,15 +5260,15 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_nc_u32_e32 v4, 0x200, v4 +; GFX12-TRUE16-NEXT: v_add_nc_u32_e32 v6, 0x200, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s1, exec_lo ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 3, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, -4, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 3, v6 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v6 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, -4, v6 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v6, v7, 0xffff -; GFX12-TRUE16-NEXT: v_not_b32_e32 v9, v6 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v7, v4, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v11, v7 ; GFX12-TRUE16-NEXT: .LBB15_1: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 ; GFX12-TRUE16-NEXT: v_readfirstlane_b32 s5, v1 @@ -5278,39 +5282,38 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: buffer_load_b32 v6, v8, s[4:7], null offen +; GFX12-TRUE16-NEXT: buffer_load_b32 v7, v10, s[4:7], null offen ; GFX12-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB15_1 ; GFX12-TRUE16-NEXT: ; %bb.2: ; GFX12-TRUE16-NEXT: s_mov_b32 exec_lo, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 ; GFX12-TRUE16-NEXT: s_mov_b32 s1, 0 ; GFX12-TRUE16-NEXT: .LBB15_3: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: ; Child Loop BB15_4 Depth 2 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v7, v6 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v6, v4, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.h, v5.l ; GFX12-TRUE16-NEXT: s_mov_b32 s2, exec_lo ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v4, v4, v10 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v6, v6, v8 +; GFX12-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v6 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v4, v5, v11, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v12, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v8.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, v4, v6 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v7, v5 -; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v9, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v5 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v5, v6 +; GFX12-TRUE16-NEXT: v_and_or_b32 v6, v7, v11, v6 +; GFX12-TRUE16-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6 ; GFX12-TRUE16-NEXT: .LBB15_4: ; Parent Loop BB15_3 Depth=1 ; GFX12-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2 ; GFX12-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 @@ -5325,14 +5328,14 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v8, s[4:7], null offen th:TH_ATOMIC_RETURN +; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[8:9], v10, s[4:7], null offen th:TH_ATOMIC_RETURN ; GFX12-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB15_4 ; GFX12-TRUE16-NEXT: ; %bb.5: ; in Loop: Header=BB15_3 Depth=1 ; GFX12-TRUE16-NEXT: s_mov_b32 exec_lo, s2 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v8, v7 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v7, v8 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_or_b32 s1, vcc_lo, s1 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -5340,7 +5343,7 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall__amd ; GFX12-TRUE16-NEXT: s_cbranch_execnz .LBB15_3 ; GFX12-TRUE16-NEXT: ; %bb.6: ; %atomicrmw.end ; GFX12-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s1 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v7, v4 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v4, v8 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -5516,16 +5519,16 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v4, 0x200, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, 0x200, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s1, 0 ; GFX11-TRUE16-NEXT: s_mov_b32 s2, exec_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 3, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, -4, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 3, v6 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v6 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, -4, v6 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v6, v7, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v9, v6 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v7, v4, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v11, v7 ; GFX11-TRUE16-NEXT: .LBB15_1: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s5, v1 @@ -5537,39 +5540,38 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-NEXT: s_and_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 -; GFX11-TRUE16-NEXT: buffer_load_b32 v6, v8, s[4:7], 0 offen +; GFX11-TRUE16-NEXT: buffer_load_b32 v7, v10, s[4:7], 0 offen ; GFX11-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB15_1 ; GFX11-TRUE16-NEXT: ; %bb.2: ; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v5 ; GFX11-TRUE16-NEXT: s_set_inst_prefetch_distance 0x1 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB15_3: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: ; Child Loop BB15_4 Depth 2 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v7, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v6, v4, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v5.l ; GFX11-TRUE16-NEXT: s_mov_b32 s2, exec_lo ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v4, v4, v10 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v5, v11, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h +; GFX11-TRUE16-NEXT: v_min_f32_e32 v6, v6, v8 +; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v12, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, v4, v6 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v7, v5 -; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v9, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v5 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v5, v6 +; GFX11-TRUE16-NEXT: v_and_or_b32 v6, v7, v11, v6 +; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6 ; GFX11-TRUE16-NEXT: .LBB15_4: ; Parent Loop BB15_3 Depth=1 ; GFX11-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2 ; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s4, v0 @@ -5583,14 +5585,14 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v8, s[4:7], 0 offen glc +; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[8:9], v10, s[4:7], 0 offen glc ; GFX11-TRUE16-NEXT: s_xor_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB15_4 ; GFX11-TRUE16-NEXT: ; %bb.5: ; in Loop: Header=BB15_3 Depth=1 ; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s2 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v6 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v8, v7 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v7, v8 ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv ; GFX11-TRUE16-NEXT: s_or_b32 s1, vcc_lo, s1 @@ -5600,7 +5602,7 @@ define bfloat @buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall__amd ; GFX11-TRUE16-NEXT: ; %bb.6: ; %atomicrmw.end ; GFX11-TRUE16-NEXT: s_set_inst_prefetch_distance 0x2 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v7, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v0, v4, v8 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: buffer_fat_ptr_agent_atomic_fmin_ret_bf16__offset__waterfall__amdgpu_no_fine_grained_memory: diff --git a/llvm/test/CodeGen/AMDGPU/calling-conventions.ll b/llvm/test/CodeGen/AMDGPU/calling-conventions.ll index a1aef8d..da49140 100644 --- a/llvm/test/CodeGen/AMDGPU/calling-conventions.ll +++ b/llvm/test/CodeGen/AMDGPU/calling-conventions.ll @@ -2748,100 +2748,101 @@ define amdgpu_cs void @amdgpu_cs_v32i1(<32 x i1> %arg0) { ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, v10.l, 1 ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 1, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, v8.l, 1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 3, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, v6.l, 1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 2, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 1, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v8.l, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, v4.l, 1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 3, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 2, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, v6.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, v6.l, 1 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 1, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, v4.l, 1 ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, v2.l, 1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 3, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.h, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, v0.h, 3 ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 1, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, v0.l, 1 -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, v28.l, 1 -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, v26.l, 1 -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, v3.l, 15 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 1, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 2, v1.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 2, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 1, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 3, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, v24.l, 1 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.h, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, v0.l, 3 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 1, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v5.l, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, v22.l, 1 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, v20.l, 1 -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, v18.l, 1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 1, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, v16.l, 1 ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, v14.l, 1 ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 1, v13.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, v12.l, 1 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, v30.l, 1 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 3, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 2, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v6.l, v5.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 3, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 2, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 3, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 2, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 3, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 3, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 2, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v8.l, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 3, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 2, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v4.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 3, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 2, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, v4.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, v5.h, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, v7.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v12.h, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, v12.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, v26.l, 1 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 1, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, v24.l, 1 +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, v22.l, 1 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 1, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, v20.l, 1 +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, v18.l, 1 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 1, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, v16.l, 1 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 3, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 2, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v12.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.h, v10.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, v5.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v7.l, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, v0.h, 3 ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.h, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, v3.l, 3 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v6.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, v0.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v30.l, v30.l, 1 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 1, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v28.l, v28.l, 1 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 3, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 2, v25.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 3, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 2, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v20.l, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 3, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 2, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v12.h, v10.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, v8.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 3, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 2, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v28.l, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v25.h +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v23.l, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, v16.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, v15.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, v1.l, 15 ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 4, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, v0.l, 15 -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.h, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, v3.l, 15 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 4, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, v4.l, 15 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 12, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v30.h, v28.h +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, v24.h, 3 +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v24.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v14.h, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v16.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 12, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 12, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v24.h, v28.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, v20.h, 15 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 4, v2.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, v1.h, 15 +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 12, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.l, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: global_store_b32 v[0:1], v0, off ; GFX11-TRUE16-NEXT: s_endpgm ; diff --git a/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll b/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll index 745e047..86e890b 100644 --- a/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll +++ b/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll @@ -1771,33 +1771,35 @@ define amdgpu_kernel void @load_v4i8_to_v4f32_2_uses(ptr addrspace(1) noalias %o ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x34 ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, 0 :: v_dual_and_b32 v0, 0x3ff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 2, v0 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: global_load_b32 v4, v0, s[0:1] ; GFX11-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v4.l, 9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 9 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff00, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff00, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff00, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff00, v4.h ; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte3_e32 v3, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte2_e32 v2, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.h, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 9 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x900, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x900, v0.h +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.h ; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte1_e32 v1, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v2 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x900, v0.l +; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte2_e32 v2, v4 ; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte0_e32 v0, v4 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v6, v7 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v6 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: global_store_b128 v5, v[0:3], s[0:1] diff --git a/llvm/test/CodeGen/AMDGPU/dagcombine-fmul-sel.ll b/llvm/test/CodeGen/AMDGPU/dagcombine-fmul-sel.ll index 738bad7..f26b720 100644 --- a/llvm/test/CodeGen/AMDGPU/dagcombine-fmul-sel.ll +++ b/llvm/test/CodeGen/AMDGPU/dagcombine-fmul-sel.ll @@ -2811,20 +2811,20 @@ define bfloat @fmul_select_bf16_test1(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0x3f80 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0x4000, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0x4000, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -2852,20 +2852,20 @@ define bfloat @fmul_select_bf16_test1(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-GISEL-TRUE16: ; %bb.0: ; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0x3f80 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0x4000, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0x4000, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -2945,20 +2945,20 @@ define bfloat @fmul_select_bf16_test2(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0x3f80 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0x3f00, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0x3f00, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -2986,20 +2986,20 @@ define bfloat @fmul_select_bf16_test2(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-GISEL-TRUE16: ; %bb.0: ; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0x3f80 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0x3f00, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0x3f00, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -3105,34 +3105,34 @@ define <2 x bfloat> @fmul_select_v2bf16_test3(<2 x bfloat> %x, <2 x i32> %bool.a ; GFX11-SDAG-TRUE16-LABEL: fmul_select_v2bf16_test3: ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0x3f80 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v3 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e64 s0, v2, v4 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, 0x4000, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.l, v5.l, 0x4000, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0x3f80 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e64 s0, v1, v3 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 +; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v2.l, 0x4000, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, 0x4000, s0 ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_dual_mul_f32 v0, v0, v2 :: v_dual_mul_f32 v1, v3, v1 -; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v2, v1, 16, 1 -; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff -; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v2, v4, v3 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v3 +; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 +; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-SDAG-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v1, v0 +; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v1 ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-SDAG-FAKE16-LABEL: fmul_select_v2bf16_test3: @@ -3170,34 +3170,34 @@ define <2 x bfloat> @fmul_select_v2bf16_test3(<2 x bfloat> %x, <2 x i32> %bool.a ; GFX11-GISEL-TRUE16-LABEL: fmul_select_v2bf16_test3: ; GFX11-GISEL-TRUE16: ; %bb.0: ; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0x3f80 -; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v3 -; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e64 s0, v2, v4 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, 0x4000, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v2.l, v5.l, 0x4000, s0 +; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0x3f80 +; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e64 s0, v1, v3 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-GISEL-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 +; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v3.h, v2.l, 0x4000, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, 0x4000, s0 ; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX11-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_dual_mul_f32 v0, v0, v2 :: v_dual_mul_f32 v1, v3, v1 -; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v2, v1, 16, 1 -; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 -; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff -; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v2, v4, v3 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v3 +; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 +; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-GISEL-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v1, v0 +; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v1 ; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-GISEL-FAKE16-LABEL: fmul_select_v2bf16_test3: @@ -3314,34 +3314,34 @@ define <2 x bfloat> @fmul_select_v2bf16_test4(<2 x bfloat> %x, <2 x i32> %bool.a ; GFX11-SDAG-TRUE16-LABEL: fmul_select_v2bf16_test4: ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0x3f80 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v3 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e64 s0, v2, v4 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, 0x3f00, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.l, v5.l, 0x3f00, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0x3f80 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e64 s0, v1, v3 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 +; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v2.l, 0x3f00, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, 0x3f00, s0 ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_dual_mul_f32 v0, v0, v2 :: v_dual_mul_f32 v1, v3, v1 -; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v2, v1, 16, 1 -; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff -; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v2, v4, v3 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v3 +; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 +; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-SDAG-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v1, v0 +; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v1 ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-SDAG-FAKE16-LABEL: fmul_select_v2bf16_test4: @@ -3379,34 +3379,34 @@ define <2 x bfloat> @fmul_select_v2bf16_test4(<2 x bfloat> %x, <2 x i32> %bool.a ; GFX11-GISEL-TRUE16-LABEL: fmul_select_v2bf16_test4: ; GFX11-GISEL-TRUE16: ; %bb.0: ; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0x3f80 -; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v3 -; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e64 s0, v2, v4 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, 0x3f00, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v2.l, v5.l, 0x3f00, s0 +; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v4 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0x3f80 +; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e64 s0, v1, v3 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-GISEL-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 +; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v3.h, v2.l, 0x3f00, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, 0x3f00, s0 ; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX11-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_dual_mul_f32 v0, v0, v2 :: v_dual_mul_f32 v1, v3, v1 -; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v2, v1, 16, 1 -; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 -; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff -; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v2, v4, v3 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v3 +; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 +; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX11-GISEL-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v1, v0 +; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v1 ; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-GISEL-FAKE16-LABEL: fmul_select_v2bf16_test4: @@ -3498,20 +3498,20 @@ define bfloat @fmul_select_bf16_test5(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0x4100 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0x4000, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0x4000, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -3539,20 +3539,20 @@ define bfloat @fmul_select_bf16_test5(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-GISEL-TRUE16: ; %bb.0: ; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0x4100 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0x4000, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0x4000, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -3634,20 +3634,20 @@ define bfloat @fmul_select_bf16_test6(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0x4040 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0xc100, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0xc100, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -3675,20 +3675,20 @@ define bfloat @fmul_select_bf16_test6(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-GISEL-TRUE16: ; %bb.0: ; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0x4040 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0xc100, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0xc100, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -3769,20 +3769,20 @@ define bfloat @fmul_select_bf16_test7(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0xc080 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0x4100, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0x4100, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -3810,20 +3810,20 @@ define bfloat @fmul_select_bf16_test7(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-GISEL-TRUE16: ; %bb.0: ; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0xc080 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0x4100, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0x4100, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -3902,12 +3902,13 @@ define bfloat @fmul_select_bf16_test8(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-SDAG-TRUE16-LABEL: fmul_select_bf16_test8: ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, 0, 0x8000, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v3.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, 0, 0x8000, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v3, v0 ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 @@ -3940,12 +3941,13 @@ define bfloat @fmul_select_bf16_test8(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-GISEL-TRUE16-LABEL: fmul_select_bf16_test8: ; GFX11-GISEL-TRUE16: ; %bb.0: ; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, 0, 0x8000, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v3.l +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v0.h, 0, 0x8000, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v3, v0 ; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 @@ -4033,20 +4035,20 @@ define bfloat @fmul_select_bf16_test9(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0xc200 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0xc180, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0xc180, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -4074,20 +4076,20 @@ define bfloat @fmul_select_bf16_test9(bfloat %x, i32 %bool.arg1, i32 %bool.arg2) ; GFX11-GISEL-TRUE16: ; %bb.0: ; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0xc200 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0xc180, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0xc180, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -4169,20 +4171,20 @@ define bfloat @fmul_select_bf16_test10_sel_log2val_pos65_pos56(bfloat %x, i32 %b ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0xdb80 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0xe000, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0xe000, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -4210,20 +4212,20 @@ define bfloat @fmul_select_bf16_test10_sel_log2val_pos65_pos56(bfloat %x, i32 %b ; GFX11-GISEL-TRUE16: ; %bb.0: ; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0xdb80 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0xe000, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0xe000, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -4305,20 +4307,20 @@ define bfloat @fmul_select_bf16_test11_sel_log2val_neg22_pos25(bfloat %x, i32 %b ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0x4c00 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0x3480, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0x3480, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -4346,20 +4348,20 @@ define bfloat @fmul_select_bf16_test11_sel_log2val_neg22_pos25(bfloat %x, i32 %b ; GFX11-GISEL-TRUE16: ; %bb.0: ; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0x4c00 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v1.l, v3.l, 0x3480, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v0, v1 +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v4.l +; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, 0x3480, vcc_lo +; GFX11-GISEL-TRUE16-NEXT: v_mul_f32_e32 v0, v4, v0 +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-GISEL-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-GISEL-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX11-GISEL-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/divergence-driven-buildvector.ll b/llvm/test/CodeGen/AMDGPU/divergence-driven-buildvector.ll index 45fe2d0..85e56a2 100644 --- a/llvm/test/CodeGen/AMDGPU/divergence-driven-buildvector.ll +++ b/llvm/test/CodeGen/AMDGPU/divergence-driven-buildvector.ll @@ -77,11 +77,20 @@ define i32 @divergent_vec_0_i16(i16 %a) { ; GFX906-NEXT: v_lshlrev_b32_e32 v0, 16, v0 ; GFX906-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: divergent_vec_0_i16: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11-TRUE16-LABEL: divergent_vec_0_i16: +; GFX11-TRUE16: ; %bb.0: +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, v1 +; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11-FAKE16-LABEL: divergent_vec_0_i16: +; GFX11-FAKE16: ; %bb.0: +; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31] %tmp = insertelement <2 x i16> poison, i16 0, i32 0 %vec = insertelement <2 x i16> %tmp, i16 %a, i32 1 %val = bitcast <2 x i16> %vec to i32 @@ -160,11 +169,20 @@ define i32 @divergent_vec_i16_0(i16 %a) { ; GFX906-NEXT: v_and_b32_e32 v0, 0xffff, v0 ; GFX906-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: divergent_vec_i16_0: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11-TRUE16-LABEL: divergent_vec_i16_0: +; GFX11-TRUE16: ; %bb.0: +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, v1 +; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11-FAKE16-LABEL: divergent_vec_i16_0: +; GFX11-FAKE16: ; %bb.0: +; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 +; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31] %tmp = insertelement <2 x i16> poison, i16 %a, i32 0 %vec = insertelement <2 x i16> %tmp, i16 0, i32 1 %val = bitcast <2 x i16> %vec to i32 @@ -243,11 +261,20 @@ define float @divergent_vec_f16_0(half %a) { ; GFX906-NEXT: v_and_b32_e32 v0, 0xffff, v0 ; GFX906-NEXT: s_setpc_b64 s[30:31] ; -; GFX11-LABEL: divergent_vec_f16_0: -; GFX11: ; %bb.0: -; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-NEXT: s_setpc_b64 s[30:31] +; GFX11-TRUE16-LABEL: divergent_vec_f16_0: +; GFX11-TRUE16: ; %bb.0: +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, v1 +; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX11-FAKE16-LABEL: divergent_vec_f16_0: +; GFX11-FAKE16: ; %bb.0: +; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 +; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31] %tmp = insertelement <2 x half> poison, half %a, i32 0 %vec = insertelement <2 x half> %tmp, half 0.0, i32 1 %val = bitcast <2 x half> %vec to float diff --git a/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll b/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll index 6cd4399..d8f81db 100644 --- a/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll +++ b/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll @@ -624,31 +624,30 @@ define amdgpu_kernel void @v_fabs_fold_self_v2bf16(ptr addrspace(1) %out, ptr ad ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0x3ff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 2, v0 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-TRUE16-NEXT: global_load_b32 v2, v0, s[2:3] +; GFX11-TRUE16-NEXT: global_load_b32 v0, v0, s[2:3] ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0x7fff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0x7fff, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_dual_mul_f32 v0, v0, v3 :: v_dual_lshlrev_b32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_mul_f32_e32 v1, v1, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0x7fff, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_dual_mul_f32 v2, v1, v2 :: v_dual_lshlrev_b32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0x7fff, v0.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mul_f32_e32 v0, v1, v3 +; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v0 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, 0 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo @@ -809,35 +808,35 @@ define amdgpu_kernel void @v_fabs_fold_v2bf16(ptr addrspace(1) %out, ptr addrspa ; GFX11-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0x3ff, v0 ; GFX11-TRUE16-NEXT: s_load_b32 s4, s[4:5], 0x10 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 2, v0 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: global_load_b32 v0, v0, s[2:3] -; GFX11-TRUE16-NEXT: s_lshl_b32 s2, s4, 16 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0x7fff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0x7fff, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_dual_mul_f32 v0, s2, v0 :: v_dual_lshlrev_b32 v1, 16, v1 ; GFX11-TRUE16-NEXT: s_and_b32 s2, s4, 0xffff0000 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1) -; GFX11-TRUE16-NEXT: v_mul_f32_e32 v1, s2, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0x7fff, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mul_f32_e32 v2, s2, v1 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0x7fff, v0.l +; GFX11-TRUE16-NEXT: s_lshl_b32 s2, s4, 16 +; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mul_f32_e32 v0, s2, v1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 +; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v0 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v1 ; GFX11-TRUE16-NEXT: global_store_b32 v2, v0, s[0:1] ; GFX11-TRUE16-NEXT: s_endpgm @@ -988,34 +987,36 @@ define amdgpu_kernel void @v_extract_fabs_fold_v2bf16(ptr addrspace(1) %in) #0 { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0x3ff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 2, v0 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: global_load_b32 v0, v0, s[0:1] ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0x7fff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0x7fff, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_dual_mul_f32 v0, 4.0, v0 :: v_dual_lshlrev_b32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 2.0, v1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v0 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0x7fff, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0x7fff, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_mul_f32_e32 v2, 4.0, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v2 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 2.0, v1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v2, 16, 1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1 +; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo -; GFX11-TRUE16-NEXT: global_store_d16_hi_b16 v[0:1], v0, off dlc -; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo ; GFX11-TRUE16-NEXT: global_store_d16_hi_b16 v[0:1], v1, off dlc ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 +; GFX11-TRUE16-NEXT: global_store_d16_hi_b16 v[0:1], v0, off dlc +; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 ; GFX11-TRUE16-NEXT: s_endpgm ; ; GFX11-FAKE16-LABEL: v_extract_fabs_fold_v2bf16: diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll index 8581e4d03..8c7d5cf 100644 --- a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll +++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll @@ -11974,7 +11974,7 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -11988,20 +11988,22 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -12121,7 +12123,7 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-LABEL: flat_agent_atomic_fadd_ret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -12136,19 +12138,21 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -12425,34 +12429,34 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB47_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -12578,12 +12582,11 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -12594,19 +12597,21 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -12891,34 +12896,34 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB48_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13045,12 +13050,11 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -13061,19 +13065,21 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13355,45 +13361,45 @@ define void @flat_agent_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB49_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -13505,45 +13511,46 @@ define void @flat_agent_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX11-TRUE16-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB49_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -13806,45 +13813,45 @@ define void @flat_agent_atomic_fadd_noret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB50_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -13957,45 +13964,46 @@ define void @flat_agent_atomic_fadd_noret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX11-TRUE16-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b_neg__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB50_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -14259,27 +14267,28 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__offset12b_pos__align4__amdgpu_no ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB51_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 ; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 @@ -14379,27 +14388,28 @@ define bfloat @flat_agent_atomic_fadd_ret_bf16__offset12b_pos__align4__amdgpu_no ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB51_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 ; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) @@ -14629,32 +14639,33 @@ define void @flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos__amdgpu_no ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] offset:2046 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB52_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v2, v2, v4 -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -14744,33 +14755,34 @@ define void @flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos__amdgpu_no ; GFX11-TRUE16-LABEL: flat_agent_atomic_fadd_noret_bf16__offset12b__align4_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] offset:2046 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB52_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -14988,7 +15000,7 @@ define void @flat_agent_atomic_fadd_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -15002,18 +15014,20 @@ define void @flat_agent_atomic_fadd_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -15130,7 +15144,7 @@ define void @flat_agent_atomic_fadd_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-LABEL: flat_agent_atomic_fadd_noret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -15145,17 +15159,19 @@ define void @flat_agent_atomic_fadd_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -15424,34 +15440,34 @@ define bfloat @flat_system_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB54_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -15579,12 +15595,11 @@ define bfloat @flat_system_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -15595,19 +15610,21 @@ define bfloat @flat_system_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -15891,46 +15908,46 @@ define void @flat_system_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB55_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SYS +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SYS -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -16043,45 +16060,46 @@ define void @flat_system_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX11-TRUE16-LABEL: flat_system_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB55_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll index 883063b..56ad91d 100644 --- a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll +++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll @@ -9836,7 +9836,7 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -9850,20 +9850,22 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9983,7 +9985,7 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-LABEL: flat_agent_atomic_fmax_ret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -9998,19 +10000,21 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10288,34 +10292,34 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB37_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10441,12 +10445,11 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -10457,19 +10460,21 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10755,34 +10760,34 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB38_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10909,12 +10914,11 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -10925,19 +10929,21 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -11220,7 +11226,7 @@ define void @flat_agent_atomic_fmax_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -11234,18 +11240,20 @@ define void @flat_agent_atomic_fmax_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -11362,7 +11370,7 @@ define void @flat_agent_atomic_fmax_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-LABEL: flat_agent_atomic_fmax_noret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -11377,17 +11385,19 @@ define void @flat_agent_atomic_fmax_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -11654,45 +11664,45 @@ define void @flat_agent_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB40_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -11804,45 +11814,46 @@ define void @flat_agent_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX11-TRUE16-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB40_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -12106,45 +12117,45 @@ define void @flat_agent_atomic_fmax_noret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -12257,45 +12268,46 @@ define void @flat_agent_atomic_fmax_noret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX11-TRUE16-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b_neg__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -12560,27 +12572,28 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__offset12b_pos__align4__amdgpu_no ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB42_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 ; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 @@ -12680,27 +12693,28 @@ define bfloat @flat_agent_atomic_fmax_ret_bf16__offset12b_pos__align4__amdgpu_no ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB42_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 ; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) @@ -12931,32 +12945,33 @@ define void @flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos__amdgpu_no ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] offset:2046 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB43_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v2, v2, v4 -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -13046,33 +13061,34 @@ define void @flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos__amdgpu_no ; GFX11-TRUE16-LABEL: flat_agent_atomic_fmax_noret_bf16__offset12b__align4_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] offset:2046 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB43_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_max_f32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -13294,34 +13310,34 @@ define bfloat @flat_system_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB44_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13449,12 +13465,11 @@ define bfloat @flat_system_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -13465,19 +13480,21 @@ define bfloat @flat_system_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13762,46 +13779,46 @@ define void @flat_system_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB45_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SYS +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SYS -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -13914,45 +13931,46 @@ define void @flat_system_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX11-TRUE16-LABEL: flat_system_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB45_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll index c603421..f0083bd 100644 --- a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll +++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll @@ -9836,7 +9836,7 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -9850,20 +9850,22 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9983,7 +9985,7 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-LABEL: flat_agent_atomic_fmin_ret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -9998,19 +10000,21 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10288,34 +10292,34 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB37_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10441,12 +10445,11 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -10457,19 +10460,21 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10755,34 +10760,34 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB38_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10909,12 +10914,11 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -10925,19 +10929,21 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -11220,7 +11226,7 @@ define void @flat_agent_atomic_fmin_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -11234,18 +11240,20 @@ define void @flat_agent_atomic_fmin_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -11362,7 +11370,7 @@ define void @flat_agent_atomic_fmin_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-LABEL: flat_agent_atomic_fmin_noret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -11377,17 +11385,19 @@ define void @flat_agent_atomic_fmin_noret_bf16__amdgpu_no_fine_grained_memory(pt ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -11654,45 +11664,45 @@ define void @flat_agent_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB40_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -11804,45 +11814,46 @@ define void @flat_agent_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine_gr ; GFX11-TRUE16-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB40_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -12106,45 +12117,45 @@ define void @flat_agent_atomic_fmin_noret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -12257,45 +12268,46 @@ define void @flat_agent_atomic_fmin_noret_bf16__offset12b_neg__amdgpu_no_fine_gr ; GFX11-TRUE16-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b_neg__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -12560,27 +12572,28 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__offset12b_pos__align4__amdgpu_no ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB42_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 ; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 @@ -12680,27 +12693,28 @@ define bfloat @flat_agent_atomic_fmin_ret_bf16__offset12b_pos__align4__amdgpu_no ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB42_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 ; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) @@ -12931,32 +12945,33 @@ define void @flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos__amdgpu_no ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] offset:2046 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB43_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v2, v2, v4 -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -13046,33 +13061,34 @@ define void @flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos__amdgpu_no ; GFX11-TRUE16-LABEL: flat_agent_atomic_fmin_noret_bf16__offset12b__align4_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] offset:2046 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB43_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_min_f32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -13294,34 +13310,34 @@ define bfloat @flat_system_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB44_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13449,12 +13465,11 @@ define bfloat @flat_system_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -13465,19 +13480,21 @@ define bfloat @flat_system_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13762,46 +13779,46 @@ define void @flat_system_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB45_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SYS +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SYS -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -13914,45 +13931,46 @@ define void @flat_system_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine_g ; GFX11-TRUE16-LABEL: flat_system_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB45_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll index c987effe..3ee0bb2 100644 --- a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll +++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll @@ -9419,7 +9419,7 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16(ptr %ptr, bfloat %val) #0 { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -9433,20 +9433,22 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16(ptr %ptr, bfloat %val) #0 { ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9566,7 +9568,7 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16(ptr %ptr, bfloat %val) #0 { ; GFX11-TRUE16-LABEL: flat_agent_atomic_fsub_ret_bf16: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -9581,19 +9583,21 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16(ptr %ptr, bfloat %val) #0 { ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9870,34 +9874,34 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat % ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB33_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10023,12 +10027,11 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat % ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -10039,19 +10042,21 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat % ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10336,34 +10341,34 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr %ptr, bfloat % ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB34_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10490,12 +10495,11 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr %ptr, bfloat % ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -10506,19 +10510,21 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr %ptr, bfloat % ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10800,7 +10806,7 @@ define void @flat_agent_atomic_fsub_noret_bf16(ptr %ptr, bfloat %val) #0 { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -10814,18 +10820,20 @@ define void @flat_agent_atomic_fsub_noret_bf16(ptr %ptr, bfloat %val) #0 { ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -10942,7 +10950,7 @@ define void @flat_agent_atomic_fsub_noret_bf16(ptr %ptr, bfloat %val) #0 { ; GFX11-TRUE16-LABEL: flat_agent_atomic_fsub_noret_bf16: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -10957,17 +10965,19 @@ define void @flat_agent_atomic_fsub_noret_bf16(ptr %ptr, bfloat %val) #0 { ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -11233,45 +11243,45 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b_pos(ptr %ptr, bfloat % ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB36_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -11383,45 +11393,46 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b_pos(ptr %ptr, bfloat % ; GFX11-TRUE16-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_pos: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB36_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -11684,45 +11695,45 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b_neg(ptr %ptr, bfloat % ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB37_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -11835,45 +11846,46 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b_neg(ptr %ptr, bfloat % ; GFX11-TRUE16-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b_neg: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB37_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -12137,27 +12149,28 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4(ptr %ptr, ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB38_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 ; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 @@ -12257,27 +12270,28 @@ define bfloat @flat_agent_atomic_fsub_ret_bf16__offset12b_pos__align4(ptr %ptr, ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB38_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 ; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) @@ -12507,32 +12521,33 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos(ptr %ptr, ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] offset:2046 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB39_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v4 -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -12622,33 +12637,34 @@ define void @flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos(ptr %ptr, ; GFX11-TRUE16-LABEL: flat_agent_atomic_fsub_noret_bf16__offset12b__align4_pos: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] offset:2046 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB39_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:2046 glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -12869,34 +12885,34 @@ define bfloat @flat_system_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB40_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13024,12 +13040,11 @@ define bfloat @flat_system_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: flat_load_b32 v5, v[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -13040,19 +13055,21 @@ define bfloat @flat_system_atomic_fsub_ret_bf16__offset12b_pos(ptr %ptr, bfloat ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13336,46 +13353,46 @@ define void @flat_system_atomic_fsub_noret_bf16__offset12b_pos(ptr %ptr, bfloat ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SYS +; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SYS -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -13488,45 +13505,46 @@ define void @flat_system_atomic_fsub_noret_bf16__offset12b_pos(ptr %ptr, bfloat ; GFX11-TRUE16-LABEL: flat_system_atomic_fsub_noret_bf16__offset12b_pos: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: flat_load_b32 v3, v[0:1] -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: flat_load_b32 v4, v[0:1] +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc +; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 diff --git a/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll b/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll index 2f08931..8784352 100644 --- a/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll +++ b/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll @@ -1872,63 +1872,48 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b, ; GFX11-SDAG-TRUE16-LABEL: v_max3_bf16_maximumnum_maximumnum__v_v_v_0: ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.h +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.h, v1.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v1, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.h, v1.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.h +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, v2.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v1, v1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v3, v4 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.l, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v3.l -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v0.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.l ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.h, v0.l, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v3.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.h +; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v3, v1 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v3.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.h +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v3.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, vcc_lo ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v2, v3 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo ; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-SDAG-FAKE16-LABEL: v_max3_bf16_maximumnum_maximumnum__v_v_v_0: @@ -1985,74 +1970,59 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b, ; GFX12-SDAG-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3) ; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v1, v1 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v3, v4 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.h ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.l, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v3.l +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.h, v1.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v1, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.h, v1.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.h ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 ; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, v2.l +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v1, v1 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v2, v3 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v0.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.l +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.h, v0.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v3.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.h +; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v3, v1 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v3.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.h +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v3.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, vcc_lo ; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) ; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX12-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-SDAG-FAKE16-LABEL: v_max3_bf16_maximumnum_maximumnum__v_v_v_0: @@ -2505,114 +2475,88 @@ define <2 x bfloat> @v_max3_v2bf16_maximumnum_maximumnum__v_v_v_0(<2 x bfloat> % ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v0 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v1 +; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v6, v6 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v0.h, v1.h, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.l, s1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v8.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v7 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v6, v8 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.l, v4.l, v3.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v6.l, v1.l, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v7 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v8 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v4.l, s1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v7 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v0.h, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v6, v6 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.l +; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v0.h, v1.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v5.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v5.h +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.h, v5.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v5.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v5.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.l, v1.l, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v4.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.h ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v4, v5 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v5.h +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.h, v4.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.h ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v2.h, v1.l, s1 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s2 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v2.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v4.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v2 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v5.h, s2 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.l, v2.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v2.h, v5.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.h +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.h, v5.h, s0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v5.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v1, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v5.l +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v0.l, v2.l, s0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v5.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.h +; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v1 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.h +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v5.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.h +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v5.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v4, v6 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v7 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.l, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v1.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v6 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v1 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v1.h, s2 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-SDAG-FAKE16-LABEL: v_max3_v2bf16_maximumnum_maximumnum__v_v_v_0: @@ -2710,123 +2654,103 @@ define <2 x bfloat> @v_max3_v2bf16_maximumnum_maximumnum__v_v_v_0(<2 x bfloat> % ; GFX12-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v0 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v0 -; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v1 +; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 ; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v6, v6 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.l +; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v0 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v0.h, v1.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v0.h, v1.h, vcc_lo ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.l, s1 -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v8.l, v1.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v7 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v6, v8 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v5.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v5.h +; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.l, v4.l, v3.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.h, v5.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v5.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v6.l, v1.l, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v5.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.l, v1.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v4.h, s1 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.h +; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v4, v5 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v5.h ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v7 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v8 -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v4.l, s1 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v7 -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v0.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.h, v4.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.h +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v4.h, s1 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v6, v6 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v2 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v5.h, s2 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s0 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v2.h, v1.l, s1 -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s2 -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v2.l +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 ; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v4, v6 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v7 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.l, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v1.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.l, v2.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.l +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v2.h, v5.h, s1 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.h +; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v6 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v1 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.h, v5.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v5.h, s1 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v1, v1 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v5.l +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v0.l, v2.l, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v5.h, s1 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.h +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v1 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.h +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v5.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.h +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v5.h, s1 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, s0 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v1.h, s2 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 ; GFX12-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-SDAG-FAKE16-LABEL: v_max3_v2bf16_maximumnum_maximumnum__v_v_v_0: diff --git a/llvm/test/CodeGen/AMDGPU/fmed3.bf16.ll b/llvm/test/CodeGen/AMDGPU/fmed3.bf16.ll index 0742ac7..bc85dc2 100644 --- a/llvm/test/CodeGen/AMDGPU/fmed3.bf16.ll +++ b/llvm/test/CodeGen/AMDGPU/fmed3.bf16.ll @@ -69,21 +69,16 @@ define bfloat @v_test_fmed3_r_i_i_bf16_minimumnum_maximumnum(bfloat %a) #1 { ; GFX11-SDAG-TRUE16-LABEL: v_test_fmed3_r_i_i_bf16_minimumnum_maximumnum: ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_cmp_o_f32_e32 vcc_lo, v1, v1 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, 0x4000, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, 0x4000, v0.l, vcc_lo ; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, 2.0, v1 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, 0x4000, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, 0x4000, v1.h, vcc_lo ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 4.0, v1 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, 0x4080, v0.l, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, 0x4080, v1.h, vcc_lo ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] %max = call bfloat @llvm.maximumnum.bf16(bfloat %a, bfloat 2.0) %med = call bfloat @llvm.minimumnum.bf16(bfloat %max, bfloat 4.0) @@ -196,35 +191,26 @@ define <2 x bfloat> @v_test_fmed3_r_i_i_v2bf16_minimumnum_maximumnum(<2 x bfloat ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff0000, v0 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) ; GFX11-SDAG-TRUE16-NEXT: v_cmp_o_f32_e32 vcc_lo, v1, v1 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_o_f32_e64 s0, v2, v2 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, 0x4000, v0.h, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, 0x4000, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v0 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.h, 0x4000, v0.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_o_f32_e64 s0, v1, v1 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, 2.0, v2 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, 0x4000, v2.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.h, 0x4000, v0.l, s0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4) ; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, 2.0, v2 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, 2.0, v3 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, 0x4000, v1.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, 0x4000, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, 0x4000, v2.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 4.0, v2 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, 4.0, v3 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l ; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, 0x4080, v1.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, 4.0, v2 ; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, 0x4080, v0.l, s0 ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] %max = call <2 x bfloat> @llvm.maximumnum.v2bf16(<2 x bfloat> %a, <2 x bfloat> splat (bfloat 2.0)) diff --git a/llvm/test/CodeGen/AMDGPU/fmin3-minimumnum.ll b/llvm/test/CodeGen/AMDGPU/fmin3-minimumnum.ll index 969c6c3..7b2d793 100644 --- a/llvm/test/CodeGen/AMDGPU/fmin3-minimumnum.ll +++ b/llvm/test/CodeGen/AMDGPU/fmin3-minimumnum.ll @@ -1874,63 +1874,48 @@ define bfloat @v_min3_bf16_minimumnum_minimumnum__v_v_v_0(bfloat %a, bfloat %b, ; GFX11-SDAG-TRUE16-LABEL: v_min3_bf16_minimumnum_minimumnum__v_v_v_0: ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.h +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.h, v1.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v1, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.h, v1.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.h +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, v2.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v1, v1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v3, v4 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.l, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v3.l -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v0.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.l ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.h, v0.l, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v3.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.h +; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v3, v1 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v3.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.h +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v3.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, vcc_lo ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v2, v3 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo ; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-SDAG-FAKE16-LABEL: v_min3_bf16_minimumnum_minimumnum__v_v_v_0: @@ -1987,74 +1972,59 @@ define bfloat @v_min3_bf16_minimumnum_minimumnum__v_v_v_0(bfloat %a, bfloat %b, ; GFX12-SDAG-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3) ; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v1, v1 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v1.l +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v3, v4 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.h ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.l, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v3.l +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.h, v1.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v1, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.h, v1.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.h ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 ; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, v2.l +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v1, v1 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v2, v3 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v0.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.l +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.h, v0.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v3.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.h +; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v3, v1 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v3.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.h +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v3.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, vcc_lo ; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) ; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX12-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-SDAG-FAKE16-LABEL: v_min3_bf16_minimumnum_minimumnum__v_v_v_0: @@ -2510,114 +2480,88 @@ define <2 x bfloat> @v_min3_v2bf16_minimumnum_minimumnum__v_v_v_0(<2 x bfloat> % ; GFX11-SDAG-TRUE16: ; %bb.0: ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v0 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v0 -; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v1 +; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v6, v6 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v0.h, v1.h, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.l, s1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v8.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v7 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v6, v8 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.l, v4.l, v3.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v6.l, v1.l, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v7 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v8 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v4.l, s1 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v7 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v0.h, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v6, v6 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.l +; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v0.h, v1.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v5.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v5.h +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.h, v5.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v5.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v5.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.l, v1.l, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v4.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.h ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v4, v5 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v5.h +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.h, v4.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.h ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v2.h, v1.l, s1 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s2 -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.l -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v2.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v4.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v2 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v5.h, s2 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.l, v2.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.l +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v2.h, v5.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.h +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v3 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.h, v5.h, s0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v5.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v1, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v5.l +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v0.l, v2.l, s0 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v5.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.h +; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v1 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.h +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v5.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.h +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v5.h, s1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v4, v6 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v7 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.l, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v1.l, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v6 -; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v1 -; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 -; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo -; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, s0 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v1.h, s2 +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo +; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 ; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-SDAG-FAKE16-LABEL: v_min3_v2bf16_minimumnum_minimumnum__v_v_v_0: @@ -2715,123 +2659,103 @@ define <2 x bfloat> @v_min3_v2bf16_minimumnum_minimumnum__v_v_v_0(<2 x bfloat> % ; GFX12-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v0 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v0 -; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v1 +; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 ; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v6, v6 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.l +; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v0 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v0.h, v1.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v0.h, v1.h, vcc_lo ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.l, s1 -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v8.l, v1.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v7 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v6, v8 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v5.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v5.h +; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.l, v4.l, v3.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.h, v5.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v5.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v6.l, v1.l, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v5.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.l, v1.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v4.h, s1 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.h +; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v4, v5 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v5.h ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v7 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v8 -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v4.l, s1 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v7 -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v0.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.h, v4.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.h +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v4.h, s1 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v6, v6 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v2 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v5.h, s2 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, vcc_lo -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s0 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v2.h, v1.l, s1 -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s2 -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.l -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.l -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v2.l +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 ; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v4, v6 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v7 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.l, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v1.l, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.l, v2.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.l +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.h, v2.h, v5.h, s1 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.h +; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v3 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v6 -; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v1 -; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.h, v5.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v5.h, s1 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v3.l +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v1, v1 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v5.l +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX12-SDAG-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v5.h, v0.l, v2.l, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v5.h, s1 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.h +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v1 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.h +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v5.h, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.h +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v5.h, s1 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v5.h, v0.l +; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.h, s0 +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v1.h, s2 +; GFX12-SDAG-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo +; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo ; GFX12-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 +; GFX12-SDAG-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 ; GFX12-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-SDAG-FAKE16-LABEL: v_min3_v2bf16_minimumnum_minimumnum__v_v_v_0: diff --git a/llvm/test/CodeGen/AMDGPU/function-args.ll b/llvm/test/CodeGen/AMDGPU/function-args.ll index a901d7f..f8ff8ef 100644 --- a/llvm/test/CodeGen/AMDGPU/function-args.ll +++ b/llvm/test/CodeGen/AMDGPU/function-args.ll @@ -1109,18 +1109,19 @@ define void @void_func_v4i8(<4 x i8> %arg0) #0 { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0 ; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h ; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1 -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: buffer_store_b32 v0, off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -1191,19 +1192,20 @@ define void @void_func_v5i8(<5 x i8> %arg0) #0 { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 4 ; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h ; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1 -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l ; GFX11-TRUE16-NEXT: buffer_store_b8 v4, off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: buffer_store_b32 v0, off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -1289,23 +1291,24 @@ define void @void_func_v8i8(<8 x i8> %arg0) #0 { ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v0.h, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0 -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v1.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v0.h, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5 ; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 -; GFX11-TRUE16-NEXT: buffer_store_b64 v[0:1], off, s[0:3], 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v4 +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v0.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v3 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v0, v4 +; GFX11-TRUE16-NEXT: buffer_store_b64 v[1:2], off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: void_func_v8i8: @@ -1419,46 +1422,47 @@ define void @void_func_v16i8(<16 x i8> %arg0) #0 { ; GFX11-TRUE16-LABEL: void_func_v16i8: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v8.h, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v9.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, 0 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v13 ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v8.h, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v10.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v11 +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v1.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v5 +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v6 ; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v4, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.l, v0.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v2 ; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1 -; GFX11-TRUE16-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v0, v12 +; GFX11-TRUE16-NEXT: buffer_store_b128 v[6:9], off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: void_func_v16i8: @@ -1654,85 +1658,84 @@ define void @void_func_v32i8(<32 x i8> %arg0) #0 { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: scratch_load_d16_u8 v31, off, s32 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v2.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v9.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v7.h, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v5.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13 ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v7.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v28.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v3.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v13, v32 +; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v5.h, v4.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v12, v32 +; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v6.l, v7.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v6.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v12.l, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v3.h, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v5.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v7.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v8.h, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v10.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v11.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v13.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v14.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v15.l, v14.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v17 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v19 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v6 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v11.h, v11.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v5 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v13, v32 +; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v8.h, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v0.h, v6.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v0.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v13, v14 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v4, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v8 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v32 +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.h, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v6 ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 16 ; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000 ; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v31.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v9.l, v0.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v31.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v9.l, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v23.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v32 +; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v10.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v32 +; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v4.h, v5.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v9, v10 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v1, v0 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v17, v18 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v12, v19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v11, v32 +; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v4.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v9, v32 ; GFX11-TRUE16-NEXT: buffer_store_b128 v[4:7], off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0 ; GFX11-TRUE16-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 diff --git a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll index dd7f183..facc91a 100644 --- a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll +++ b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll @@ -4896,22 +4896,23 @@ define amdgpu_gfx void @test_call_external_void_func_v4i8_ret() #0 { ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 16, v0 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, 24, v0 ; GFX11-TRUE16-NEXT: s_swappc_b64 s[30:31], s[0:1] -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l ; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v42, 1 ; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v42, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 ; GFX11-TRUE16-NEXT: s_mov_b32 s32, s33 ; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: global_store_b32 v[40:41], v0, off ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s33 @@ -5156,18 +5157,22 @@ define amdgpu_gfx void @test_call_external_void_func_v5i8_ret() #0 { ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v6 ; GFX11-TRUE16-NEXT: s_swappc_b64 s[30:31], s[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v1.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, 4 :: v_dual_lshlrev_b32 v3, 16, v1 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, 0 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v0.h +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, 4 ; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v42, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_and_b32 v2, 0xffff, v2 +; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v42, 0 +; GFX11-TRUE16-NEXT: s_mov_b32 s32, s33 +; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v3 ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: global_store_b8 v[0:1], v4, off @@ -5175,9 +5180,6 @@ define amdgpu_gfx void @test_call_external_void_func_v5i8_ret() #0 { ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s33 ; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s33 offset:4 -; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v42, 0 -; GFX11-TRUE16-NEXT: s_mov_b32 s32, s33 -; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2 ; GFX11-TRUE16-NEXT: s_or_saveexec_b32 s1, -1 ; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s33 offset:8 ; 4-byte Folded Reload ; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s1 @@ -5440,36 +5442,35 @@ define amdgpu_gfx void @test_call_external_void_func_v8i8_ret() #0 { ; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v8 ; GFX11-TRUE16-NEXT: s_swappc_b64 s[30:31], s[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v5.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v3.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4 -; GFX11-TRUE16-NEXT: global_store_b64 v[40:41], v[0:1], off -; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s33 -; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s33 offset:4 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v1.h, v1.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v4 +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v5 +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v1.l, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v4 ; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v42, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v0, v5 ; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v42, 0 ; GFX11-TRUE16-NEXT: s_mov_b32 s32, s33 ; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2 +; GFX11-TRUE16-NEXT: global_store_b64 v[40:41], v[1:2], off +; GFX11-TRUE16-NEXT: s_clause 0x1 +; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s33 +; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s33 offset:4 ; GFX11-TRUE16-NEXT: s_or_saveexec_b32 s1, -1 ; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s33 offset:8 ; 4-byte Folded Reload ; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s1 @@ -5911,81 +5912,83 @@ define amdgpu_gfx void @test_call_external_void_func_v32i8_ret() #0 { ; GFX11-TRUE16-NEXT: s_swappc_b64 s[30:31], s[0:1] ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v13.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v15.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, 0 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l ; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v3.h, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v5.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v5 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v7, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v4, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v3.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v12, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v2.h, v1.h -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v9 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v21.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v12 +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v0.h, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v3.h, v1.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v9, v13 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8 +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v1.l, v0.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v6.l, v0.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v9 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v10, v8 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v0 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v1, v2 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v3, v7 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v6, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v1.l, v0.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v1.l +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v7, v13 +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v6.l, v0.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v6, v13 +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v1.l, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v10 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v0, v13 ; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: global_store_b128 v[42:43], v[7:10], off -; GFX11-TRUE16-NEXT: global_store_b128 v[40:41], v[3:6], off +; GFX11-TRUE16-NEXT: global_store_b128 v[42:43], v[6:9], off +; GFX11-TRUE16-NEXT: global_store_b128 v[40:41], v[2:5], off ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s33 ; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s33 offset:4 diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll index 100a560..1f74fbd 100644 --- a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll +++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll @@ -12320,7 +12320,7 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -12334,20 +12334,22 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -12467,7 +12469,7 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-LABEL: global_agent_atomic_fadd_ret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -12482,19 +12484,21 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -12821,34 +12825,34 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB55_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -12974,12 +12978,11 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -12990,19 +12993,21 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13339,34 +13344,34 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB56_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13493,12 +13498,11 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -13509,19 +13513,21 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13855,7 +13861,7 @@ define void @global_agent_atomic_fadd_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -13869,18 +13875,20 @@ define void @global_agent_atomic_fadd_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -13997,7 +14005,7 @@ define void @global_agent_atomic_fadd_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-LABEL: global_agent_atomic_fadd_noret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -14012,17 +14020,19 @@ define void @global_agent_atomic_fadd_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -14337,45 +14347,45 @@ define void @global_agent_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB58_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -14487,45 +14497,46 @@ define void @global_agent_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX11-TRUE16-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB58_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -14838,45 +14849,45 @@ define void @global_agent_atomic_fadd_noret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB59_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -14989,45 +15000,46 @@ define void @global_agent_atomic_fadd_noret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX11-TRUE16-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b_neg__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB59_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -15341,27 +15353,28 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4__amdgpu_ ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB60_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 ; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 @@ -15461,27 +15474,28 @@ define bfloat @global_agent_atomic_fadd_ret_bf16__offset12b_pos__align4__amdgpu_ ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB60_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 ; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) @@ -15750,32 +15764,33 @@ define void @global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos__amdgpu_ ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off offset:2046 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB61_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v2, v2, v4 -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -15865,33 +15880,34 @@ define void @global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos__amdgpu_ ; GFX11-TRUE16-LABEL: global_agent_atomic_fadd_noret_bf16__offset12b__align4_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off offset:2046 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB61_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -16149,34 +16165,34 @@ define bfloat @global_system_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB62_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -16304,12 +16320,11 @@ define bfloat @global_system_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -16320,19 +16335,21 @@ define bfloat @global_system_atomic_fadd_ret_bf16__offset12b_pos__amdgpu_no_fine ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -16668,46 +16685,46 @@ define void @global_system_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB63_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SYS -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -16820,45 +16837,46 @@ define void @global_system_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine ; GFX11-TRUE16-LABEL: global_system_atomic_fadd_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB63_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll index faa3ee6..faa74fe 100644 --- a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll +++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll @@ -8741,7 +8741,7 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -8755,20 +8755,22 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -8888,7 +8890,7 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-LABEL: global_agent_atomic_fmax_ret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -8903,19 +8905,21 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9244,34 +9248,34 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB37_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9397,12 +9401,11 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -9413,19 +9416,21 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9764,34 +9769,34 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB38_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9918,12 +9923,11 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -9934,19 +9938,21 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10282,7 +10288,7 @@ define void @global_agent_atomic_fmax_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -10296,18 +10302,20 @@ define void @global_agent_atomic_fmax_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -10424,7 +10432,7 @@ define void @global_agent_atomic_fmax_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-LABEL: global_agent_atomic_fmax_noret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -10439,17 +10447,19 @@ define void @global_agent_atomic_fmax_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -10766,45 +10776,45 @@ define void @global_agent_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB40_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -10916,45 +10926,46 @@ define void @global_agent_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX11-TRUE16-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB40_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -11269,45 +11280,45 @@ define void @global_agent_atomic_fmax_noret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -11420,45 +11431,46 @@ define void @global_agent_atomic_fmax_noret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX11-TRUE16-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b_neg__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -11774,27 +11786,28 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4__amdgpu_ ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB42_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 ; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 @@ -11894,27 +11907,28 @@ define bfloat @global_agent_atomic_fmax_ret_bf16__offset12b_pos__align4__amdgpu_ ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB42_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 ; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) @@ -12185,32 +12199,33 @@ define void @global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos__amdgpu_ ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off offset:2046 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB43_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v2, v2, v4 -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -12300,33 +12315,34 @@ define void @global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos__amdgpu_ ; GFX11-TRUE16-LABEL: global_agent_atomic_fmax_noret_bf16__offset12b__align4_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off offset:2046 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB43_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_max_f32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -12586,34 +12602,34 @@ define bfloat @global_system_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB44_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -12741,12 +12757,11 @@ define bfloat @global_system_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -12757,19 +12772,21 @@ define bfloat @global_system_atomic_fmax_ret_bf16__offset12b_pos__amdgpu_no_fine ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_max_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13107,46 +13124,46 @@ define void @global_system_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB45_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_max_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SYS -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -13259,45 +13276,46 @@ define void @global_system_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine ; GFX11-TRUE16-LABEL: global_system_atomic_fmax_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB45_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_max_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_max_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll index cb66f85..a46b012 100644 --- a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll +++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll @@ -8741,7 +8741,7 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -8755,20 +8755,22 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -8888,7 +8890,7 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-LABEL: global_agent_atomic_fmin_ret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -8903,19 +8905,21 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9244,34 +9248,34 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB37_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9397,12 +9401,11 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -9413,19 +9416,21 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9764,34 +9769,34 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB38_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9918,12 +9923,11 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -9934,19 +9938,21 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10282,7 +10288,7 @@ define void @global_agent_atomic_fmin_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -10296,18 +10302,20 @@ define void @global_agent_atomic_fmin_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -10424,7 +10432,7 @@ define void @global_agent_atomic_fmin_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-LABEL: global_agent_atomic_fmin_noret_bf16__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -10439,17 +10447,19 @@ define void @global_agent_atomic_fmin_noret_bf16__amdgpu_no_fine_grained_memory( ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -10766,45 +10776,45 @@ define void @global_agent_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB40_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -10916,45 +10926,46 @@ define void @global_agent_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine_ ; GFX11-TRUE16-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB40_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -11269,45 +11280,45 @@ define void @global_agent_atomic_fmin_noret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -11420,45 +11431,46 @@ define void @global_agent_atomic_fmin_noret_bf16__offset12b_neg__amdgpu_no_fine_ ; GFX11-TRUE16-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b_neg__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -11774,27 +11786,28 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4__amdgpu_ ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB42_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 ; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 @@ -11894,27 +11907,28 @@ define bfloat @global_agent_atomic_fmin_ret_bf16__offset12b_pos__align4__amdgpu_ ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB42_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 ; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) @@ -12185,32 +12199,33 @@ define void @global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos__amdgpu_ ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off offset:2046 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB43_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v2, v2, v4 -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -12300,33 +12315,34 @@ define void @global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos__amdgpu_ ; GFX11-TRUE16-LABEL: global_agent_atomic_fmin_noret_bf16__offset12b__align4_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off offset:2046 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB43_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_min_f32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -12586,34 +12602,34 @@ define bfloat @global_system_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB44_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -12741,12 +12757,11 @@ define bfloat @global_system_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -12757,19 +12772,21 @@ define bfloat @global_system_atomic_fmin_ret_bf16__offset12b_pos__amdgpu_no_fine ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_min_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13107,46 +13124,46 @@ define void @global_system_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB45_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_min_num_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SYS -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -13259,45 +13276,46 @@ define void @global_system_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine ; GFX11-TRUE16-LABEL: global_system_atomic_fmin_noret_bf16__offset12b_pos__amdgpu_no_fine_grained_memory: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB45_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_min_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_min_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll index f869b57..053efdc 100644 --- a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll +++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll @@ -9266,7 +9266,7 @@ define bfloat @global_agent_atomic_fsub_ret_bf16(ptr addrspace(1) %ptr, bfloat % ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -9280,20 +9280,22 @@ define bfloat @global_agent_atomic_fsub_ret_bf16(ptr addrspace(1) %ptr, bfloat % ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9413,7 +9415,7 @@ define bfloat @global_agent_atomic_fsub_ret_bf16(ptr addrspace(1) %ptr, bfloat % ; GFX11-TRUE16-LABEL: global_agent_atomic_fsub_ret_bf16: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -9428,19 +9430,21 @@ define bfloat @global_agent_atomic_fsub_ret_bf16(ptr addrspace(1) %ptr, bfloat % ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9767,34 +9771,34 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1) ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB33_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -9920,12 +9924,11 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1) ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -9936,19 +9939,21 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1) ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10285,34 +10290,34 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr addrspace(1) ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB34_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10439,12 +10444,11 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr addrspace(1) ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -10455,19 +10459,21 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_neg(ptr addrspace(1) ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -10801,7 +10807,7 @@ define void @global_agent_atomic_fsub_noret_bf16(ptr addrspace(1) %ptr, bfloat % ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -10815,18 +10821,20 @@ define void @global_agent_atomic_fsub_noret_bf16(ptr addrspace(1) %ptr, bfloat % ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -10943,7 +10951,7 @@ define void @global_agent_atomic_fsub_noret_bf16(ptr addrspace(1) %ptr, bfloat % ; GFX11-TRUE16-LABEL: global_agent_atomic_fsub_noret_bf16: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_lshlrev_b32 v2, 16, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 @@ -10958,17 +10966,19 @@ define void @global_agent_atomic_fsub_noret_bf16(ptr addrspace(1) %ptr, bfloat % ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 @@ -11283,45 +11293,45 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b_pos(ptr addrspace(1) ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB36_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -11433,45 +11443,46 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b_pos(ptr addrspace(1) ; GFX11-TRUE16-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_pos: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB36_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -11784,45 +11795,45 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b_neg(ptr addrspace(1) ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB37_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -11935,45 +11946,46 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b_neg(ptr addrspace(1) ; GFX11-TRUE16-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b_neg: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0xfffff800, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0xfffff800, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, -1, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB37_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -12287,27 +12299,28 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4(ptr addr ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB38_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 ; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 @@ -12407,27 +12420,28 @@ define bfloat @global_agent_atomic_fsub_ret_bf16__offset12b_pos__align4(ptr addr ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB38_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v5 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 ; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) @@ -12696,32 +12710,33 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos(ptr addr ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off offset:2046 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: .LBB39_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v4 -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_DEV -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -12811,33 +12826,34 @@ define void @global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos(ptr addr ; GFX11-TRUE16-LABEL: global_agent_atomic_fsub_noret_bf16__offset12b__align4_pos: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off offset:2046 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off offset:2046 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB39_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v3.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.h -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v3, v5 +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off offset:2046 glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -13095,34 +13111,34 @@ define bfloat @global_system_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1 ; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX12-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v4, v3, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_not_b32_e32 v4, v4 ; GFX12-TRUE16-NEXT: .LBB40_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13250,12 +13266,11 @@ define bfloat @global_system_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1 ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 +; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 -; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 ; GFX11-TRUE16-NEXT: global_load_b32 v5, v[0:1], off ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) @@ -13266,19 +13281,21 @@ define bfloat @global_system_atomic_fsub_ret_bf16__offset12b_pos(ptr addrspace(1 ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v7 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) @@ -13614,46 +13631,46 @@ define void @global_system_atomic_fsub_noret_bf16__offset12b_pos(ptr addrspace(1 ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 +; GFX12-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX12-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX12-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX12-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS +; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_SYS ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SYS -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -13766,45 +13783,46 @@ define void @global_system_atomic_fsub_noret_bf16__offset12b_pos(ptr addrspace(1 ; GFX11-TRUE16-LABEL: global_system_atomic_fsub_noret_bf16__offset12b_pos: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_co_u32 v4, vcc_lo, 0x7fe, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add_co_u32 v3, vcc_lo, 0x7fe, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 3, v4 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: global_load_b32 v3, v[0:1], off -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 3, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, -4, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 3, v3 +; GFX11-TRUE16-NEXT: global_load_b32 v4, v[0:1], off +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v5, v4, 0xffff -; GFX11-TRUE16-NEXT: v_not_b32_e32 v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v5, 0xffff +; GFX11-TRUE16-NEXT: v_not_b32_e32 v6, v3 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB41_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, v4, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v2, v2, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v2, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v8, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, v4, v7 -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, v3, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v7 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v2, v[0:1], v[2:3], off glc +; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl1_inv ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 diff --git a/llvm/test/CodeGen/AMDGPU/global-saddr-load.ll b/llvm/test/CodeGen/AMDGPU/global-saddr-load.ll index d588f0e..723e3ef 100644 --- a/llvm/test/CodeGen/AMDGPU/global-saddr-load.ll +++ b/llvm/test/CodeGen/AMDGPU/global-saddr-load.ll @@ -4007,6 +4007,8 @@ define amdgpu_ps <2 x half> @global_load_saddr_i16_d16lo_zero_hi(ptr addrspace(1 ; GFX12-GISEL-TRUE16-NEXT: global_load_d16_b16 v0, v0, s[2:3] ; GFX12-GISEL-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 +; GFX12-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-GISEL-TRUE16-NEXT: v_lshl_or_b32 v0, 0, 16, v0 ; GFX12-GISEL-TRUE16-NEXT: ; return to shader part epilog ; ; GFX12-GISEL-FAKE16-LABEL: global_load_saddr_i16_d16lo_zero_hi: @@ -4053,6 +4055,8 @@ define amdgpu_ps <2 x half> @global_load_saddr_i16_d16lo_zero_hi_immneg128(ptr a ; GFX12-GISEL-TRUE16-NEXT: global_load_d16_b16 v0, v0, s[2:3] offset:-128 ; GFX12-GISEL-TRUE16-NEXT: s_wait_loadcnt 0x0 ; GFX12-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0 +; GFX12-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-GISEL-TRUE16-NEXT: v_lshl_or_b32 v0, 0, 16, v0 ; GFX12-GISEL-TRUE16-NEXT: ; return to shader part epilog ; ; GFX12-GISEL-FAKE16-LABEL: global_load_saddr_i16_d16lo_zero_hi_immneg128: @@ -4411,7 +4415,7 @@ define amdgpu_ps <2 x half> @global_load_saddr_i16_d16hi_zero_hi(ptr addrspace(1 ; GFX12-GISEL-TRUE16: ; %bb.0: ; GFX12-GISEL-TRUE16-NEXT: global_load_d16_b16 v0, v0, s[2:3] ; GFX12-GISEL-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX12-GISEL-TRUE16-NEXT: v_lshl_or_b32 v0, v0, 16, 0 ; GFX12-GISEL-TRUE16-NEXT: ; return to shader part epilog ; ; GFX12-GISEL-FAKE16-LABEL: global_load_saddr_i16_d16hi_zero_hi: @@ -4457,7 +4461,7 @@ define amdgpu_ps <2 x half> @global_load_saddr_i16_d16hi_zero_hi_immneg128(ptr a ; GFX12-GISEL-TRUE16: ; %bb.0: ; GFX12-GISEL-TRUE16-NEXT: global_load_d16_b16 v0, v0, s[2:3] offset:-128 ; GFX12-GISEL-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX12-GISEL-TRUE16-NEXT: v_lshl_or_b32 v0, v0, 16, 0 ; GFX12-GISEL-TRUE16-NEXT: ; return to shader part epilog ; ; GFX12-GISEL-FAKE16-LABEL: global_load_saddr_i16_d16hi_zero_hi_immneg128: @@ -4882,7 +4886,7 @@ define amdgpu_ps void @global_addr_64bit_lsr_iv(ptr addrspace(1) inreg %arg) { ; GFX12-GISEL-NEXT: v_mov_b32_e32 v2, s0 ; GFX12-GISEL-NEXT: .LBB132_1: ; %bb3 ; GFX12-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 -; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) ; GFX12-GISEL-NEXT: v_add_co_u32 v4, vcc, v0, v2 ; GFX12-GISEL-NEXT: s_wait_alu 0xfffd ; GFX12-GISEL-NEXT: v_add_co_ci_u32_e64 v5, null, v1, v3, vcc @@ -5002,7 +5006,7 @@ define amdgpu_ps void @global_addr_64bit_lsr_iv_multiload(ptr addrspace(1) inreg ; GFX12-GISEL-NEXT: v_mov_b32_e32 v2, s0 ; GFX12-GISEL-NEXT: .LBB133_1: ; %bb3 ; GFX12-GISEL-NEXT: ; =>This Inner Loop Header: Depth=1 -; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) ; GFX12-GISEL-NEXT: v_add_co_u32 v4, vcc, v0, v2 ; GFX12-GISEL-NEXT: s_wait_alu 0xfffd ; GFX12-GISEL-NEXT: v_add_co_ci_u32_e64 v5, null, v1, v3, vcc diff --git a/llvm/test/CodeGen/AMDGPU/idot4u.ll b/llvm/test/CodeGen/AMDGPU/idot4u.ll index 5e50288..7ebd692 100644 --- a/llvm/test/CodeGen/AMDGPU/idot4u.ll +++ b/llvm/test/CodeGen/AMDGPU/idot4u.ll @@ -2724,31 +2724,31 @@ define amdgpu_kernel void @udot4_acc8_vecMul(ptr addrspace(1) %src1, ; GFX11-DL-TRUE16-NEXT: global_load_b32 v4, v0, s[2:3] ; GFX11-DL-TRUE16-NEXT: global_load_d16_u8 v0, v5, s[4:5] ; GFX11-DL-TRUE16-NEXT: s_waitcnt vmcnt(2) -; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 24, v3 +; GFX11-DL-TRUE16-NEXT: v_lshrrev_b16 v0.h, 8, v3.l ; GFX11-DL-TRUE16-NEXT: s_waitcnt vmcnt(1) +; GFX11-DL-TRUE16-NEXT: v_lshrrev_b16 v1.l, 8, v4.l +; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 24, v3 ; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v6, 24, v4 -; GFX11-DL-TRUE16-NEXT: v_lshrrev_b16 v0.h, 8, v3.l -; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v1.l, v3.h, v4.h -; GFX11-DL-TRUE16-NEXT: v_lshrrev_b16 v1.h, 8, v4.l ; GFX11-DL-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-DL-TRUE16-NEXT: v_mad_u16 v0.l, v3.l, v4.l, v0.l -; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v2.l, v2.l, v6.l -; GFX11-DL-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v0.h, v0.h, v1.h -; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-DL-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v2.l +; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v0.h, v0.h, v1.l +; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v1.l, v3.h, v4.h +; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v1.h, v2.l, v6.l +; GFX11-DL-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 +; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-DL-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v0.h -; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-DL-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v6.l +; GFX11-DL-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v1.l +; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-DL-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.h ; GFX11-DL-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-DL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-DL-TRUE16-NEXT: v_or_b32_e32 v1, v2, v1 -; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-DL-TRUE16-NEXT: v_or_b16 v6.h, v0.h, v1.l ; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1 -; GFX11-DL-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l -; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v6 -; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-DL-TRUE16-NEXT: v_or_b32_e32 v2, v2, v6 +; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 8, v2 +; GFX11-DL-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v2.l +; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-DL-TRUE16-NEXT: v_mad_u16 v0.l, v3.h, v4.h, v0.l ; GFX11-DL-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l ; GFX11-DL-TRUE16-NEXT: global_store_b8 v5, v0, s[4:5] diff --git a/llvm/test/CodeGen/AMDGPU/isel-amdgpu-cs-chain-preserve-cc.ll b/llvm/test/CodeGen/AMDGPU/isel-amdgpu-cs-chain-preserve-cc.ll index 0d33400..3261e4c 100644 --- a/llvm/test/CodeGen/AMDGPU/isel-amdgpu-cs-chain-preserve-cc.ll +++ b/llvm/test/CodeGen/AMDGPU/isel-amdgpu-cs-chain-preserve-cc.ll @@ -917,47 +917,91 @@ define amdgpu_cs_chain_preserve void @amdgpu_cs_chain_preserve_cc_half(half inre } define amdgpu_cs_chain_preserve void @amdgpu_cs_chain_cc_bfloat(bfloat inreg %a, bfloat %b) { - ; DAGISEL-GFX11-WF32-LABEL: name: amdgpu_cs_chain_cc_bfloat - ; DAGISEL-GFX11-WF32: bb.0 (%ir-block.0): - ; DAGISEL-GFX11-WF32-NEXT: liveins: $sgpr0, $vgpr8 - ; DAGISEL-GFX11-WF32-NEXT: {{ $}} - ; DAGISEL-GFX11-WF32-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr8 - ; DAGISEL-GFX11-WF32-NEXT: [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr0 - ; DAGISEL-GFX11-WF32-NEXT: [[V_LSHLREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHLREV_B32_e64 16, [[COPY]], implicit $exec - ; DAGISEL-GFX11-WF32-NEXT: [[S_LSHL_B32_:%[0-9]+]]:sreg_32 = S_LSHL_B32 [[COPY1]], 16, implicit-def dead $scc - ; DAGISEL-GFX11-WF32-NEXT: [[V_ADD_F32_e64_:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, killed [[S_LSHL_B32_]], 0, killed [[V_LSHLREV_B32_e64_]], 0, 0, implicit $mode, implicit $exec - ; DAGISEL-GFX11-WF32-NEXT: [[V_BFE_U32_e64_:%[0-9]+]]:vgpr_32 = V_BFE_U32_e64 [[V_ADD_F32_e64_]], 16, 1, implicit $exec - ; DAGISEL-GFX11-WF32-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 32767 - ; DAGISEL-GFX11-WF32-NEXT: [[V_ADD3_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD3_U32_e64 killed [[V_BFE_U32_e64_]], [[V_ADD_F32_e64_]], killed [[S_MOV_B32_]], implicit $exec - ; DAGISEL-GFX11-WF32-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 4194304 - ; DAGISEL-GFX11-WF32-NEXT: [[V_OR_B32_e64_:%[0-9]+]]:vgpr_32 = V_OR_B32_e64 [[V_ADD_F32_e64_]], killed [[S_MOV_B32_1]], implicit $exec - ; DAGISEL-GFX11-WF32-NEXT: [[V_CMP_U_F32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = nofpexcept V_CMP_U_F32_e64 0, [[V_ADD_F32_e64_]], 0, [[V_ADD_F32_e64_]], 0, implicit $mode, implicit $exec - ; DAGISEL-GFX11-WF32-NEXT: [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, killed [[V_ADD3_U32_e64_]], 0, killed [[V_OR_B32_e64_]], killed [[V_CMP_U_F32_e64_]], implicit $exec - ; DAGISEL-GFX11-WF32-NEXT: [[DEF:%[0-9]+]]:sreg_64 = IMPLICIT_DEF - ; DAGISEL-GFX11-WF32-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY [[DEF]] - ; DAGISEL-GFX11-WF32-NEXT: FLAT_STORE_SHORT_D16_HI killed [[COPY2]], killed [[V_CNDMASK_B32_e64_]], 0, 0, implicit $exec, implicit $flat_scr :: (store (s16) into `ptr poison`) - ; DAGISEL-GFX11-WF32-NEXT: S_ENDPGM 0 + ; DAGISEL-GFX11-WF32-TRUE16-LABEL: name: amdgpu_cs_chain_cc_bfloat + ; DAGISEL-GFX11-WF32-TRUE16: bb.0 (%ir-block.0): + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: liveins: $sgpr0, $vgpr8 + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: {{ $}} + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr8 + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr0 + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[V_MOV_B16_t16_e64_:%[0-9]+]]:vgpr_16 = V_MOV_B16_t16_e64 0, 0, 0, implicit $exec + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vgpr_32 = REG_SEQUENCE killed [[V_MOV_B16_t16_e64_]], %subreg.lo16, [[COPY]], %subreg.hi16 + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[S_LSHL_B32_:%[0-9]+]]:sreg_32 = S_LSHL_B32 [[COPY1]], 16, implicit-def dead $scc + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[V_ADD_F32_e64_:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, killed [[S_LSHL_B32_]], 0, killed [[REG_SEQUENCE]], 0, 0, implicit $mode, implicit $exec + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[V_BFE_U32_e64_:%[0-9]+]]:vgpr_32 = V_BFE_U32_e64 [[V_ADD_F32_e64_]], 16, 1, implicit $exec + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 32767 + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[V_ADD3_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD3_U32_e64 killed [[V_BFE_U32_e64_]], [[V_ADD_F32_e64_]], killed [[S_MOV_B32_]], implicit $exec + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 4194304 + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[V_OR_B32_e64_:%[0-9]+]]:vgpr_32 = V_OR_B32_e64 [[V_ADD_F32_e64_]], killed [[S_MOV_B32_1]], implicit $exec + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[V_CMP_U_F32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = nofpexcept V_CMP_U_F32_e64 0, [[V_ADD_F32_e64_]], 0, [[V_ADD_F32_e64_]], 0, implicit $mode, implicit $exec + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, killed [[V_ADD3_U32_e64_]], 0, killed [[V_OR_B32_e64_]], killed [[V_CMP_U_F32_e64_]], implicit $exec + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[DEF:%[0-9]+]]:sreg_64 = IMPLICIT_DEF + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY [[DEF]] + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: FLAT_STORE_SHORT_D16_HI killed [[COPY2]], killed [[V_CNDMASK_B32_e64_]], 0, 0, implicit $exec, implicit $flat_scr :: (store (s16) into `ptr poison`) + ; DAGISEL-GFX11-WF32-TRUE16-NEXT: S_ENDPGM 0 ; - ; DAGISEL-GFX11-WF64-LABEL: name: amdgpu_cs_chain_cc_bfloat - ; DAGISEL-GFX11-WF64: bb.0 (%ir-block.0): - ; DAGISEL-GFX11-WF64-NEXT: liveins: $sgpr0, $vgpr8 - ; DAGISEL-GFX11-WF64-NEXT: {{ $}} - ; DAGISEL-GFX11-WF64-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr8 - ; DAGISEL-GFX11-WF64-NEXT: [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr0 - ; DAGISEL-GFX11-WF64-NEXT: [[V_LSHLREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHLREV_B32_e64 16, [[COPY]], implicit $exec - ; DAGISEL-GFX11-WF64-NEXT: [[S_LSHL_B32_:%[0-9]+]]:sreg_32 = S_LSHL_B32 [[COPY1]], 16, implicit-def dead $scc - ; DAGISEL-GFX11-WF64-NEXT: [[V_ADD_F32_e64_:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, killed [[S_LSHL_B32_]], 0, killed [[V_LSHLREV_B32_e64_]], 0, 0, implicit $mode, implicit $exec - ; DAGISEL-GFX11-WF64-NEXT: [[V_BFE_U32_e64_:%[0-9]+]]:vgpr_32 = V_BFE_U32_e64 [[V_ADD_F32_e64_]], 16, 1, implicit $exec - ; DAGISEL-GFX11-WF64-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 32767 - ; DAGISEL-GFX11-WF64-NEXT: [[V_ADD3_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD3_U32_e64 killed [[V_BFE_U32_e64_]], [[V_ADD_F32_e64_]], killed [[S_MOV_B32_]], implicit $exec - ; DAGISEL-GFX11-WF64-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 4194304 - ; DAGISEL-GFX11-WF64-NEXT: [[V_OR_B32_e64_:%[0-9]+]]:vgpr_32 = V_OR_B32_e64 [[V_ADD_F32_e64_]], killed [[S_MOV_B32_1]], implicit $exec - ; DAGISEL-GFX11-WF64-NEXT: [[V_CMP_U_F32_e64_:%[0-9]+]]:sreg_64_xexec = nofpexcept V_CMP_U_F32_e64 0, [[V_ADD_F32_e64_]], 0, [[V_ADD_F32_e64_]], 0, implicit $mode, implicit $exec - ; DAGISEL-GFX11-WF64-NEXT: [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, killed [[V_ADD3_U32_e64_]], 0, killed [[V_OR_B32_e64_]], killed [[V_CMP_U_F32_e64_]], implicit $exec - ; DAGISEL-GFX11-WF64-NEXT: [[DEF:%[0-9]+]]:sreg_64 = IMPLICIT_DEF - ; DAGISEL-GFX11-WF64-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY [[DEF]] - ; DAGISEL-GFX11-WF64-NEXT: FLAT_STORE_SHORT_D16_HI killed [[COPY2]], killed [[V_CNDMASK_B32_e64_]], 0, 0, implicit $exec, implicit $flat_scr :: (store (s16) into `ptr poison`) - ; DAGISEL-GFX11-WF64-NEXT: S_ENDPGM 0 + ; DAGISEL-GFX11-WF32-FAKE16-LABEL: name: amdgpu_cs_chain_cc_bfloat + ; DAGISEL-GFX11-WF32-FAKE16: bb.0 (%ir-block.0): + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: liveins: $sgpr0, $vgpr8 + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: {{ $}} + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr8 + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr0 + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[V_LSHLREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHLREV_B32_e64 16, [[COPY]], implicit $exec + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[S_LSHL_B32_:%[0-9]+]]:sreg_32 = S_LSHL_B32 [[COPY1]], 16, implicit-def dead $scc + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[V_ADD_F32_e64_:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, killed [[S_LSHL_B32_]], 0, killed [[V_LSHLREV_B32_e64_]], 0, 0, implicit $mode, implicit $exec + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[V_BFE_U32_e64_:%[0-9]+]]:vgpr_32 = V_BFE_U32_e64 [[V_ADD_F32_e64_]], 16, 1, implicit $exec + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 32767 + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[V_ADD3_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD3_U32_e64 killed [[V_BFE_U32_e64_]], [[V_ADD_F32_e64_]], killed [[S_MOV_B32_]], implicit $exec + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 4194304 + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[V_OR_B32_e64_:%[0-9]+]]:vgpr_32 = V_OR_B32_e64 [[V_ADD_F32_e64_]], killed [[S_MOV_B32_1]], implicit $exec + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[V_CMP_U_F32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = nofpexcept V_CMP_U_F32_e64 0, [[V_ADD_F32_e64_]], 0, [[V_ADD_F32_e64_]], 0, implicit $mode, implicit $exec + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, killed [[V_ADD3_U32_e64_]], 0, killed [[V_OR_B32_e64_]], killed [[V_CMP_U_F32_e64_]], implicit $exec + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[DEF:%[0-9]+]]:sreg_64 = IMPLICIT_DEF + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY [[DEF]] + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: FLAT_STORE_SHORT_D16_HI killed [[COPY2]], killed [[V_CNDMASK_B32_e64_]], 0, 0, implicit $exec, implicit $flat_scr :: (store (s16) into `ptr poison`) + ; DAGISEL-GFX11-WF32-FAKE16-NEXT: S_ENDPGM 0 + ; + ; DAGISEL-GFX11-WF64-TRUE16-LABEL: name: amdgpu_cs_chain_cc_bfloat + ; DAGISEL-GFX11-WF64-TRUE16: bb.0 (%ir-block.0): + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: liveins: $sgpr0, $vgpr8 + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: {{ $}} + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr8 + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr0 + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[V_MOV_B16_t16_e64_:%[0-9]+]]:vgpr_16 = V_MOV_B16_t16_e64 0, 0, 0, implicit $exec + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vgpr_32 = REG_SEQUENCE killed [[V_MOV_B16_t16_e64_]], %subreg.lo16, [[COPY]], %subreg.hi16 + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[S_LSHL_B32_:%[0-9]+]]:sreg_32 = S_LSHL_B32 [[COPY1]], 16, implicit-def dead $scc + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[V_ADD_F32_e64_:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, killed [[S_LSHL_B32_]], 0, killed [[REG_SEQUENCE]], 0, 0, implicit $mode, implicit $exec + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[V_BFE_U32_e64_:%[0-9]+]]:vgpr_32 = V_BFE_U32_e64 [[V_ADD_F32_e64_]], 16, 1, implicit $exec + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 32767 + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[V_ADD3_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD3_U32_e64 killed [[V_BFE_U32_e64_]], [[V_ADD_F32_e64_]], killed [[S_MOV_B32_]], implicit $exec + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 4194304 + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[V_OR_B32_e64_:%[0-9]+]]:vgpr_32 = V_OR_B32_e64 [[V_ADD_F32_e64_]], killed [[S_MOV_B32_1]], implicit $exec + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[V_CMP_U_F32_e64_:%[0-9]+]]:sreg_64_xexec = nofpexcept V_CMP_U_F32_e64 0, [[V_ADD_F32_e64_]], 0, [[V_ADD_F32_e64_]], 0, implicit $mode, implicit $exec + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, killed [[V_ADD3_U32_e64_]], 0, killed [[V_OR_B32_e64_]], killed [[V_CMP_U_F32_e64_]], implicit $exec + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[DEF:%[0-9]+]]:sreg_64 = IMPLICIT_DEF + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY [[DEF]] + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: FLAT_STORE_SHORT_D16_HI killed [[COPY2]], killed [[V_CNDMASK_B32_e64_]], 0, 0, implicit $exec, implicit $flat_scr :: (store (s16) into `ptr poison`) + ; DAGISEL-GFX11-WF64-TRUE16-NEXT: S_ENDPGM 0 + ; + ; DAGISEL-GFX11-WF64-FAKE16-LABEL: name: amdgpu_cs_chain_cc_bfloat + ; DAGISEL-GFX11-WF64-FAKE16: bb.0 (%ir-block.0): + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: liveins: $sgpr0, $vgpr8 + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: {{ $}} + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr8 + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr0 + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[V_LSHLREV_B32_e64_:%[0-9]+]]:vgpr_32 = V_LSHLREV_B32_e64 16, [[COPY]], implicit $exec + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[S_LSHL_B32_:%[0-9]+]]:sreg_32 = S_LSHL_B32 [[COPY1]], 16, implicit-def dead $scc + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[V_ADD_F32_e64_:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, killed [[S_LSHL_B32_]], 0, killed [[V_LSHLREV_B32_e64_]], 0, 0, implicit $mode, implicit $exec + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[V_BFE_U32_e64_:%[0-9]+]]:vgpr_32 = V_BFE_U32_e64 [[V_ADD_F32_e64_]], 16, 1, implicit $exec + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 32767 + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[V_ADD3_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD3_U32_e64 killed [[V_BFE_U32_e64_]], [[V_ADD_F32_e64_]], killed [[S_MOV_B32_]], implicit $exec + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 4194304 + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[V_OR_B32_e64_:%[0-9]+]]:vgpr_32 = V_OR_B32_e64 [[V_ADD_F32_e64_]], killed [[S_MOV_B32_1]], implicit $exec + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[V_CMP_U_F32_e64_:%[0-9]+]]:sreg_64_xexec = nofpexcept V_CMP_U_F32_e64 0, [[V_ADD_F32_e64_]], 0, [[V_ADD_F32_e64_]], 0, implicit $mode, implicit $exec + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[V_CNDMASK_B32_e64_:%[0-9]+]]:vgpr_32 = V_CNDMASK_B32_e64 0, killed [[V_ADD3_U32_e64_]], 0, killed [[V_OR_B32_e64_]], killed [[V_CMP_U_F32_e64_]], implicit $exec + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[DEF:%[0-9]+]]:sreg_64 = IMPLICIT_DEF + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: [[COPY2:%[0-9]+]]:vreg_64 = COPY [[DEF]] + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: FLAT_STORE_SHORT_D16_HI killed [[COPY2]], killed [[V_CNDMASK_B32_e64_]], 0, 0, implicit $exec, implicit $flat_scr :: (store (s16) into `ptr poison`) + ; DAGISEL-GFX11-WF64-FAKE16-NEXT: S_ENDPGM 0 ; ; DAGISEL-GFX10-WF32-LABEL: name: amdgpu_cs_chain_cc_bfloat ; DAGISEL-GFX10-WF32: bb.0 (%ir-block.0): diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.cvt.sat.pk.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.cvt.sat.pk.ll new file mode 100644 index 0000000..3a55070 --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.cvt.sat.pk.ll @@ -0,0 +1,305 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 +; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 -mattr=+real-true16 %s -o - | FileCheck -check-prefix=SDAG-REAL16 %s +; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 -mattr=-real-true16 %s -o - | FileCheck -check-prefix=SDAG-FAKE16 %s +; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 -mattr=+real-true16 %s -o - | FileCheck -check-prefix=GISEL-REAL16 %s +; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 -mattr=-real-true16 %s -o - | FileCheck -check-prefix=GISEL-FAKE16 %s + +declare i16 @llvm.amdgcn.sat.pk4.i4.i8(i32) #0 +declare i16 @llvm.amdgcn.sat.pk4.u4.u8(i32) #0 + +define amdgpu_kernel void @sat_pk4_i4_i8_f32_v(i32 %src, ptr %out) #1 { +; SDAG-REAL16-LABEL: sat_pk4_i4_i8_f32_v: +; SDAG-REAL16: ; %bb.0: +; SDAG-REAL16-NEXT: s_clause 0x1 +; SDAG-REAL16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; SDAG-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; SDAG-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; SDAG-REAL16-NEXT: s_wait_kmcnt 0x0 +; SDAG-REAL16-NEXT: v_sat_pk4_i4_i8_e32 v0.l, s2 +; SDAG-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; SDAG-REAL16-NEXT: s_endpgm +; +; SDAG-FAKE16-LABEL: sat_pk4_i4_i8_f32_v: +; SDAG-FAKE16: ; %bb.0: +; SDAG-FAKE16-NEXT: s_clause 0x1 +; SDAG-FAKE16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; SDAG-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; SDAG-FAKE16-NEXT: v_mov_b32_e32 v0, 0 +; SDAG-FAKE16-NEXT: s_wait_kmcnt 0x0 +; SDAG-FAKE16-NEXT: v_sat_pk4_i4_i8_e32 v1, s2 +; SDAG-FAKE16-NEXT: flat_store_b16 v0, v1, s[0:1] +; SDAG-FAKE16-NEXT: s_endpgm +; +; GISEL-REAL16-LABEL: sat_pk4_i4_i8_f32_v: +; GISEL-REAL16: ; %bb.0: +; GISEL-REAL16-NEXT: s_clause 0x1 +; GISEL-REAL16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; GISEL-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; GISEL-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-REAL16-NEXT: s_wait_kmcnt 0x0 +; GISEL-REAL16-NEXT: v_sat_pk4_i4_i8_e32 v0.l, s2 +; GISEL-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-REAL16-NEXT: s_endpgm +; +; GISEL-FAKE16-LABEL: sat_pk4_i4_i8_f32_v: +; GISEL-FAKE16: ; %bb.0: +; GISEL-FAKE16-NEXT: s_clause 0x1 +; GISEL-FAKE16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; GISEL-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; GISEL-FAKE16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-FAKE16-NEXT: s_wait_kmcnt 0x0 +; GISEL-FAKE16-NEXT: v_sat_pk4_i4_i8_e32 v0, s2 +; GISEL-FAKE16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-FAKE16-NEXT: s_endpgm + %cvt = call i16 @llvm.amdgcn.sat.pk4.i4.i8(i32 %src) #0 + store i16 %cvt, ptr %out, align 2 + ret void +} + +define amdgpu_kernel void @sat_pk4_i4_i8_f32_s(i32 inreg %src, ptr %out) #1 { +; SDAG-REAL16-LABEL: sat_pk4_i4_i8_f32_s: +; SDAG-REAL16: ; %bb.1: +; SDAG-REAL16-NEXT: s_load_b32 s8, s[4:5], 0x0 +; SDAG-REAL16-NEXT: s_waitcnt lgkmcnt(0) +; SDAG-REAL16-NEXT: s_branch .LBB1_0 +; SDAG-REAL16-NEXT: .p2align 8 +; SDAG-REAL16-NEXT: ; %bb.2: +; SDAG-REAL16-NEXT: .LBB1_0: +; SDAG-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; SDAG-REAL16-NEXT: v_sat_pk4_i4_i8_e32 v0.l, s8 +; SDAG-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; SDAG-REAL16-NEXT: s_wait_kmcnt 0x0 +; SDAG-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; SDAG-REAL16-NEXT: s_endpgm +; +; SDAG-FAKE16-LABEL: sat_pk4_i4_i8_f32_s: +; SDAG-FAKE16: ; %bb.1: +; SDAG-FAKE16-NEXT: s_load_b32 s8, s[4:5], 0x0 +; SDAG-FAKE16-NEXT: s_waitcnt lgkmcnt(0) +; SDAG-FAKE16-NEXT: s_branch .LBB1_0 +; SDAG-FAKE16-NEXT: .p2align 8 +; SDAG-FAKE16-NEXT: ; %bb.2: +; SDAG-FAKE16-NEXT: .LBB1_0: +; SDAG-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; SDAG-FAKE16-NEXT: v_mov_b32_e32 v0, 0 +; SDAG-FAKE16-NEXT: v_sat_pk4_i4_i8_e32 v1, s8 +; SDAG-FAKE16-NEXT: s_wait_kmcnt 0x0 +; SDAG-FAKE16-NEXT: flat_store_b16 v0, v1, s[0:1] +; SDAG-FAKE16-NEXT: s_endpgm +; +; GISEL-REAL16-LABEL: sat_pk4_i4_i8_f32_s: +; GISEL-REAL16: ; %bb.0: +; GISEL-REAL16-NEXT: s_clause 0x1 +; GISEL-REAL16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; GISEL-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; GISEL-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-REAL16-NEXT: s_wait_kmcnt 0x0 +; GISEL-REAL16-NEXT: v_sat_pk4_i4_i8_e32 v0.l, s2 +; GISEL-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-REAL16-NEXT: s_endpgm +; +; GISEL-FAKE16-LABEL: sat_pk4_i4_i8_f32_s: +; GISEL-FAKE16: ; %bb.0: +; GISEL-FAKE16-NEXT: s_clause 0x1 +; GISEL-FAKE16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; GISEL-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; GISEL-FAKE16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-FAKE16-NEXT: s_wait_kmcnt 0x0 +; GISEL-FAKE16-NEXT: v_sat_pk4_i4_i8_e32 v0, s2 +; GISEL-FAKE16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-FAKE16-NEXT: s_endpgm + %cvt = call i16 @llvm.amdgcn.sat.pk4.i4.i8(i32 %src) #0 + store i16 %cvt, ptr %out, align 2 + ret void +} + +define amdgpu_kernel void @sat_pk4_i4_i8_f32_i(ptr %out) #1 { +; SDAG-REAL16-LABEL: sat_pk4_i4_i8_f32_i: +; SDAG-REAL16: ; %bb.0: +; SDAG-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0 +; SDAG-REAL16-NEXT: v_sat_pk4_i4_i8_e32 v0.l, 0x64 +; SDAG-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; SDAG-REAL16-NEXT: s_wait_kmcnt 0x0 +; SDAG-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; SDAG-REAL16-NEXT: s_endpgm +; +; SDAG-FAKE16-LABEL: sat_pk4_i4_i8_f32_i: +; SDAG-FAKE16: ; %bb.0: +; SDAG-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0 +; SDAG-FAKE16-NEXT: v_mov_b32_e32 v0, 0 +; SDAG-FAKE16-NEXT: v_sat_pk4_i4_i8_e32 v1, 0x64 +; SDAG-FAKE16-NEXT: s_wait_kmcnt 0x0 +; SDAG-FAKE16-NEXT: flat_store_b16 v0, v1, s[0:1] +; SDAG-FAKE16-NEXT: s_endpgm +; +; GISEL-REAL16-LABEL: sat_pk4_i4_i8_f32_i: +; GISEL-REAL16: ; %bb.0: +; GISEL-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0 +; GISEL-REAL16-NEXT: v_sat_pk4_i4_i8_e32 v0.l, 0x64 +; GISEL-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-REAL16-NEXT: s_wait_kmcnt 0x0 +; GISEL-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-REAL16-NEXT: s_endpgm +; +; GISEL-FAKE16-LABEL: sat_pk4_i4_i8_f32_i: +; GISEL-FAKE16: ; %bb.0: +; GISEL-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0 +; GISEL-FAKE16-NEXT: v_sat_pk4_i4_i8_e32 v0, 0x64 +; GISEL-FAKE16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-FAKE16-NEXT: s_wait_kmcnt 0x0 +; GISEL-FAKE16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-FAKE16-NEXT: s_endpgm + %cvt = call i16 @llvm.amdgcn.sat.pk4.i4.i8(i32 100) #0 + store i16 %cvt, ptr %out, align 2 + ret void +} + +define amdgpu_kernel void @sat_pk4_u4_u8_f32_v(i32 %src, ptr %out) #1 { +; SDAG-REAL16-LABEL: sat_pk4_u4_u8_f32_v: +; SDAG-REAL16: ; %bb.0: +; SDAG-REAL16-NEXT: s_clause 0x1 +; SDAG-REAL16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; SDAG-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; SDAG-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; SDAG-REAL16-NEXT: s_wait_kmcnt 0x0 +; SDAG-REAL16-NEXT: v_sat_pk4_u4_u8_e32 v0.l, s2 +; SDAG-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; SDAG-REAL16-NEXT: s_endpgm +; +; SDAG-FAKE16-LABEL: sat_pk4_u4_u8_f32_v: +; SDAG-FAKE16: ; %bb.0: +; SDAG-FAKE16-NEXT: s_clause 0x1 +; SDAG-FAKE16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; SDAG-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; SDAG-FAKE16-NEXT: v_mov_b32_e32 v0, 0 +; SDAG-FAKE16-NEXT: s_wait_kmcnt 0x0 +; SDAG-FAKE16-NEXT: v_sat_pk4_u4_u8_e32 v1, s2 +; SDAG-FAKE16-NEXT: flat_store_b16 v0, v1, s[0:1] +; SDAG-FAKE16-NEXT: s_endpgm +; +; GISEL-REAL16-LABEL: sat_pk4_u4_u8_f32_v: +; GISEL-REAL16: ; %bb.0: +; GISEL-REAL16-NEXT: s_clause 0x1 +; GISEL-REAL16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; GISEL-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; GISEL-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-REAL16-NEXT: s_wait_kmcnt 0x0 +; GISEL-REAL16-NEXT: v_sat_pk4_u4_u8_e32 v0.l, s2 +; GISEL-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-REAL16-NEXT: s_endpgm +; +; GISEL-FAKE16-LABEL: sat_pk4_u4_u8_f32_v: +; GISEL-FAKE16: ; %bb.0: +; GISEL-FAKE16-NEXT: s_clause 0x1 +; GISEL-FAKE16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; GISEL-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; GISEL-FAKE16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-FAKE16-NEXT: s_wait_kmcnt 0x0 +; GISEL-FAKE16-NEXT: v_sat_pk4_u4_u8_e32 v0, s2 +; GISEL-FAKE16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-FAKE16-NEXT: s_endpgm + %cvt = call i16 @llvm.amdgcn.sat.pk4.u4.u8(i32 %src) #0 + store i16 %cvt, ptr %out, align 2 + ret void +} + +define amdgpu_kernel void @sat_pk4_u4_u8_f32_s(i32 inreg %src, ptr %out) #1 { +; SDAG-REAL16-LABEL: sat_pk4_u4_u8_f32_s: +; SDAG-REAL16: ; %bb.1: +; SDAG-REAL16-NEXT: s_load_b32 s8, s[4:5], 0x0 +; SDAG-REAL16-NEXT: s_waitcnt lgkmcnt(0) +; SDAG-REAL16-NEXT: s_branch .LBB4_0 +; SDAG-REAL16-NEXT: .p2align 8 +; SDAG-REAL16-NEXT: ; %bb.2: +; SDAG-REAL16-NEXT: .LBB4_0: +; SDAG-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; SDAG-REAL16-NEXT: v_sat_pk4_u4_u8_e32 v0.l, s8 +; SDAG-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; SDAG-REAL16-NEXT: s_wait_kmcnt 0x0 +; SDAG-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; SDAG-REAL16-NEXT: s_endpgm +; +; SDAG-FAKE16-LABEL: sat_pk4_u4_u8_f32_s: +; SDAG-FAKE16: ; %bb.1: +; SDAG-FAKE16-NEXT: s_load_b32 s8, s[4:5], 0x0 +; SDAG-FAKE16-NEXT: s_waitcnt lgkmcnt(0) +; SDAG-FAKE16-NEXT: s_branch .LBB4_0 +; SDAG-FAKE16-NEXT: .p2align 8 +; SDAG-FAKE16-NEXT: ; %bb.2: +; SDAG-FAKE16-NEXT: .LBB4_0: +; SDAG-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; SDAG-FAKE16-NEXT: v_mov_b32_e32 v0, 0 +; SDAG-FAKE16-NEXT: v_sat_pk4_u4_u8_e32 v1, s8 +; SDAG-FAKE16-NEXT: s_wait_kmcnt 0x0 +; SDAG-FAKE16-NEXT: flat_store_b16 v0, v1, s[0:1] +; SDAG-FAKE16-NEXT: s_endpgm +; +; GISEL-REAL16-LABEL: sat_pk4_u4_u8_f32_s: +; GISEL-REAL16: ; %bb.0: +; GISEL-REAL16-NEXT: s_clause 0x1 +; GISEL-REAL16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; GISEL-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; GISEL-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-REAL16-NEXT: s_wait_kmcnt 0x0 +; GISEL-REAL16-NEXT: v_sat_pk4_u4_u8_e32 v0.l, s2 +; GISEL-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-REAL16-NEXT: s_endpgm +; +; GISEL-FAKE16-LABEL: sat_pk4_u4_u8_f32_s: +; GISEL-FAKE16: ; %bb.0: +; GISEL-FAKE16-NEXT: s_clause 0x1 +; GISEL-FAKE16-NEXT: s_load_b32 s2, s[4:5], 0x0 +; GISEL-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x8 +; GISEL-FAKE16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-FAKE16-NEXT: s_wait_kmcnt 0x0 +; GISEL-FAKE16-NEXT: v_sat_pk4_u4_u8_e32 v0, s2 +; GISEL-FAKE16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-FAKE16-NEXT: s_endpgm + %cvt = call i16 @llvm.amdgcn.sat.pk4.u4.u8(i32 %src) #0 + store i16 %cvt, ptr %out, align 2 + ret void +} + +define amdgpu_kernel void @sat_pk4_u4_u8_f32_i(ptr %out) #1 { +; SDAG-REAL16-LABEL: sat_pk4_u4_u8_f32_i: +; SDAG-REAL16: ; %bb.0: +; SDAG-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0 +; SDAG-REAL16-NEXT: v_sat_pk4_u4_u8_e32 v0.l, 0x64 +; SDAG-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; SDAG-REAL16-NEXT: s_wait_kmcnt 0x0 +; SDAG-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; SDAG-REAL16-NEXT: s_endpgm +; +; SDAG-FAKE16-LABEL: sat_pk4_u4_u8_f32_i: +; SDAG-FAKE16: ; %bb.0: +; SDAG-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0 +; SDAG-FAKE16-NEXT: v_mov_b32_e32 v0, 0 +; SDAG-FAKE16-NEXT: v_sat_pk4_u4_u8_e32 v1, 0x64 +; SDAG-FAKE16-NEXT: s_wait_kmcnt 0x0 +; SDAG-FAKE16-NEXT: flat_store_b16 v0, v1, s[0:1] +; SDAG-FAKE16-NEXT: s_endpgm +; +; GISEL-REAL16-LABEL: sat_pk4_u4_u8_f32_i: +; GISEL-REAL16: ; %bb.0: +; GISEL-REAL16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0 +; GISEL-REAL16-NEXT: v_sat_pk4_u4_u8_e32 v0.l, 0x64 +; GISEL-REAL16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-REAL16-NEXT: s_wait_kmcnt 0x0 +; GISEL-REAL16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-REAL16-NEXT: s_endpgm +; +; GISEL-FAKE16-LABEL: sat_pk4_u4_u8_f32_i: +; GISEL-FAKE16: ; %bb.0: +; GISEL-FAKE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x0 +; GISEL-FAKE16-NEXT: v_sat_pk4_u4_u8_e32 v0, 0x64 +; GISEL-FAKE16-NEXT: v_mov_b32_e32 v1, 0 +; GISEL-FAKE16-NEXT: s_wait_kmcnt 0x0 +; GISEL-FAKE16-NEXT: flat_store_b16 v1, v0, s[0:1] +; GISEL-FAKE16-NEXT: s_endpgm + %cvt = call i16 @llvm.amdgcn.sat.pk4.u4.u8(i32 100) #0 + store i16 %cvt, ptr %out, align 2 + ret void +} + +attributes #0 = { nounwind memory(none) } +attributes #1 = { nounwind } diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane16.swap.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane16.swap.ll index 8140866..ed6a02b 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane16.swap.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane16.swap.ll @@ -1,6 +1,8 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py -; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 < %s | FileCheck -check-prefix=GCN %s -; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 < %s | FileCheck -check-prefix=GCN %s +; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 < %s | FileCheck -check-prefix=GFX950 %s +; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx950 < %s | FileCheck -check-prefix=GFX950 %s +; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck -check-prefix=GFX1250 %s +; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck -check-prefix=GFX1250 %s ; RUN: not --crash llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -filetype=null %s 2>&1 | FileCheck -check-prefix=ERR-SDAG %s ; RUN: not llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -filetype=null %s 2>&1 | FileCheck -check-prefix=ERR-GISEL %s @@ -17,6 +19,18 @@ define { i32, i32 } @v_permlane16_swap_b32_vv(i32 %vdst_old, i32 %src0_old) { ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GCN-NEXT: v_permlane16_swap_b32_e32 v0, v1 ; GCN-NEXT: s_setpc_b64 s[30:31] +; GFX950-LABEL: v_permlane16_swap_b32_vv: +; GFX950: ; %bb.0: +; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX950-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX950-NEXT: s_setpc_b64 s[30:31] +; +; GFX1250-LABEL: v_permlane16_swap_b32_vv: +; GFX1250: ; %bb.0: +; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 +; GFX1250-NEXT: s_wait_kmcnt 0x0 +; GFX1250-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX1250-NEXT: s_set_pc_i64 s[30:31] %v = call { i32, i32 } @llvm.amdgcn.permlane16.swap(i32 %vdst_old, i32 %src0_old, i1 false, i1 false) ret { i32, i32 } %v } @@ -29,6 +43,22 @@ define { i32, i32 } @v_permlane16_swap_b32_vi(i32 %vdst_old) { ; GCN-NEXT: s_nop 1 ; GCN-NEXT: v_permlane16_swap_b32_e32 v0, v1 ; GCN-NEXT: s_setpc_b64 s[30:31] +; GFX950-LABEL: v_permlane16_swap_b32_vi: +; GFX950: ; %bb.0: +; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX950-NEXT: v_mov_b32_e32 v1, 1 +; GFX950-NEXT: s_nop 1 +; GFX950-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX950-NEXT: s_setpc_b64 s[30:31] +; +; GFX1250-LABEL: v_permlane16_swap_b32_vi: +; GFX1250: ; %bb.0: +; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 +; GFX1250-NEXT: s_wait_kmcnt 0x0 +; GFX1250-NEXT: v_mov_b32_e32 v1, 1 +; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX1250-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX1250-NEXT: s_set_pc_i64 s[30:31] %v = call { i32, i32 } @llvm.amdgcn.permlane16.swap(i32 %vdst_old, i32 1, i1 false, i1 false) ret { i32, i32 } %v } @@ -41,6 +71,22 @@ define { i32, i32 } @v_permlane16_swap_b32_vl(i32 %vdst_old) { ; GCN-NEXT: s_nop 1 ; GCN-NEXT: v_permlane16_swap_b32_e32 v0, v1 ; GCN-NEXT: s_setpc_b64 s[30:31] +; GFX950-LABEL: v_permlane16_swap_b32_vl: +; GFX950: ; %bb.0: +; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX950-NEXT: v_mov_b32_e32 v1, 0xc1d1 +; GFX950-NEXT: s_nop 1 +; GFX950-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX950-NEXT: s_setpc_b64 s[30:31] +; +; GFX1250-LABEL: v_permlane16_swap_b32_vl: +; GFX1250: ; %bb.0: +; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 +; GFX1250-NEXT: s_wait_kmcnt 0x0 +; GFX1250-NEXT: v_mov_b32_e32 v1, 0xc1d1 +; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX1250-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX1250-NEXT: s_set_pc_i64 s[30:31] %v = call { i32, i32 } @llvm.amdgcn.permlane16.swap(i32 %vdst_old, i32 49617, i1 false, i1 false) ret { i32, i32 } %v } @@ -54,6 +100,23 @@ define { i32, i32 } @v_permlane16_swap_b32_iv(i32 %src0_old) { ; GCN-NEXT: s_nop 1 ; GCN-NEXT: v_permlane16_swap_b32_e32 v0, v1 ; GCN-NEXT: s_setpc_b64 s[30:31] +; GFX950-LABEL: v_permlane16_swap_b32_iv: +; GFX950: ; %bb.0: +; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX950-NEXT: v_mov_b32_e32 v1, v0 +; GFX950-NEXT: v_mov_b32_e32 v0, 1 +; GFX950-NEXT: s_nop 1 +; GFX950-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX950-NEXT: s_setpc_b64 s[30:31] +; +; GFX1250-LABEL: v_permlane16_swap_b32_iv: +; GFX1250: ; %bb.0: +; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 +; GFX1250-NEXT: s_wait_kmcnt 0x0 +; GFX1250-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, 1 +; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX1250-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX1250-NEXT: s_set_pc_i64 s[30:31] %v = call { i32, i32 } @llvm.amdgcn.permlane16.swap(i32 1, i32 %src0_old, i1 false, i1 false) ret { i32, i32 } %v } @@ -67,6 +130,23 @@ define { i32, i32 } @v_permlane16_swap_b32_ss(i32 inreg %vdst_old, i32 inreg %sr ; GCN-NEXT: s_nop 1 ; GCN-NEXT: v_permlane16_swap_b32_e32 v0, v1 ; GCN-NEXT: s_setpc_b64 s[30:31] +; GFX950-LABEL: v_permlane16_swap_b32_ss: +; GFX950: ; %bb.0: +; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX950-NEXT: v_mov_b32_e32 v0, s0 +; GFX950-NEXT: v_mov_b32_e32 v1, s1 +; GFX950-NEXT: s_nop 1 +; GFX950-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX950-NEXT: s_setpc_b64 s[30:31] +; +; GFX1250-LABEL: v_permlane16_swap_b32_ss: +; GFX1250: ; %bb.0: +; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 +; GFX1250-NEXT: s_wait_kmcnt 0x0 +; GFX1250-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1 +; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX1250-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX1250-NEXT: s_set_pc_i64 s[30:31] %v = call { i32, i32 } @llvm.amdgcn.permlane16.swap(i32 %vdst_old, i32 %src0_old, i1 false, i1 false) ret { i32, i32 } %v } @@ -80,6 +160,23 @@ define { i32, i32 } @v_permlane16_swap_b32_sv(i32 inreg %vdst_old, i32 %src0_old ; GCN-NEXT: s_nop 1 ; GCN-NEXT: v_permlane16_swap_b32_e32 v0, v1 ; GCN-NEXT: s_setpc_b64 s[30:31] +; GFX950-LABEL: v_permlane16_swap_b32_sv: +; GFX950: ; %bb.0: +; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX950-NEXT: v_mov_b32_e32 v1, v0 +; GFX950-NEXT: v_mov_b32_e32 v0, s0 +; GFX950-NEXT: s_nop 1 +; GFX950-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX950-NEXT: s_setpc_b64 s[30:31] +; +; GFX1250-LABEL: v_permlane16_swap_b32_sv: +; GFX1250: ; %bb.0: +; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 +; GFX1250-NEXT: s_wait_kmcnt 0x0 +; GFX1250-NEXT: v_dual_mov_b32 v1, v0 :: v_dual_mov_b32 v0, s0 +; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX1250-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX1250-NEXT: s_set_pc_i64 s[30:31] %v = call { i32, i32 } @llvm.amdgcn.permlane16.swap(i32 %vdst_old, i32 %src0_old, i1 false, i1 false) ret { i32, i32 } %v } @@ -92,6 +189,22 @@ define { i32, i32 } @v_permlane16_swap_b32_vs(i32 %vdst_old, i32 inreg %src0_old ; GCN-NEXT: s_nop 1 ; GCN-NEXT: v_permlane16_swap_b32_e32 v0, v1 ; GCN-NEXT: s_setpc_b64 s[30:31] +; GFX950-LABEL: v_permlane16_swap_b32_vs: +; GFX950: ; %bb.0: +; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX950-NEXT: v_mov_b32_e32 v1, s0 +; GFX950-NEXT: s_nop 1 +; GFX950-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX950-NEXT: s_setpc_b64 s[30:31] +; +; GFX1250-LABEL: v_permlane16_swap_b32_vs: +; GFX1250: ; %bb.0: +; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 +; GFX1250-NEXT: s_wait_kmcnt 0x0 +; GFX1250-NEXT: v_mov_b32_e32 v1, s0 +; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX1250-NEXT: v_permlane16_swap_b32_e32 v0, v1 +; GFX1250-NEXT: s_set_pc_i64 s[30:31] %v = call { i32, i32 } @llvm.amdgcn.permlane16.swap(i32 %vdst_old, i32 %src0_old, i1 false, i1 false) ret { i32, i32 } %v } @@ -102,6 +215,18 @@ define { i32, i32 } @v_permlane16_swap_b32_vv_fi(i32 %vdst_old, i32 %src0_old) { ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GCN-NEXT: v_permlane16_swap_b32_e64 v0, v1 fi:1 ; GCN-NEXT: s_setpc_b64 s[30:31] +; GFX950-LABEL: v_permlane16_swap_b32_vv_fi: +; GFX950: ; %bb.0: +; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX950-NEXT: v_permlane16_swap_b32_e64 v0, v1 fi:1 +; GFX950-NEXT: s_setpc_b64 s[30:31] +; +; GFX1250-LABEL: v_permlane16_swap_b32_vv_fi: +; GFX1250: ; %bb.0: +; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 +; GFX1250-NEXT: s_wait_kmcnt 0x0 +; GFX1250-NEXT: v_permlane16_swap_b32_e64 v0, v1 fi:1 +; GFX1250-NEXT: s_set_pc_i64 s[30:31] %v = call { i32, i32 } @llvm.amdgcn.permlane16.swap(i32 %vdst_old, i32 %src0_old, i1 true, i1 false) ret { i32, i32 } %v } @@ -112,6 +237,18 @@ define { i32, i32 } @v_permlane16_swap_b32_vv_bc(i32 %vdst_old, i32 %src0_old) { ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GCN-NEXT: v_permlane16_swap_b32_e64 v0, v1 bound_ctrl:1 ; GCN-NEXT: s_setpc_b64 s[30:31] +; GFX950-LABEL: v_permlane16_swap_b32_vv_bc: +; GFX950: ; %bb.0: +; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX950-NEXT: v_permlane16_swap_b32_e64 v0, v1 bound_ctrl:1 +; GFX950-NEXT: s_setpc_b64 s[30:31] +; +; GFX1250-LABEL: v_permlane16_swap_b32_vv_bc: +; GFX1250: ; %bb.0: +; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 +; GFX1250-NEXT: s_wait_kmcnt 0x0 +; GFX1250-NEXT: v_permlane16_swap_b32_e64 v0, v1 bound_ctrl:1 +; GFX1250-NEXT: s_set_pc_i64 s[30:31] %v = call { i32, i32 } @llvm.amdgcn.permlane16.swap(i32 %vdst_old, i32 %src0_old, i1 false, i1 true) ret { i32, i32 } %v } @@ -122,6 +259,18 @@ define { i32, i32 } @v_permlane16_swap_b32_vv_fi_bc(i32 %vdst_old, i32 %src0_old ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GCN-NEXT: v_permlane16_swap_b32_e64 v0, v1 bound_ctrl:1 fi:1 ; GCN-NEXT: s_setpc_b64 s[30:31] +; GFX950-LABEL: v_permlane16_swap_b32_vv_fi_bc: +; GFX950: ; %bb.0: +; GFX950-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX950-NEXT: v_permlane16_swap_b32_e64 v0, v1 bound_ctrl:1 fi:1 +; GFX950-NEXT: s_setpc_b64 s[30:31] +; +; GFX1250-LABEL: v_permlane16_swap_b32_vv_fi_bc: +; GFX1250: ; %bb.0: +; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0 +; GFX1250-NEXT: s_wait_kmcnt 0x0 +; GFX1250-NEXT: v_permlane16_swap_b32_e64 v0, v1 bound_ctrl:1 fi:1 +; GFX1250-NEXT: s_set_pc_i64 s[30:31] %v = call { i32, i32 } @llvm.amdgcn.permlane16.swap(i32 %vdst_old, i32 %src0_old, i1 true, i1 true) ret { i32, i32 } %v } diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.prng.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.prng.ll index 2faf375..465414c 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.prng.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.prng.ll @@ -1,5 +1,7 @@ ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx950 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN %s ; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx950 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s +; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1250 < %s | FileCheck -check-prefixes=GCN %s +; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1250 < %s | FileCheck -check-prefix=GCN %s declare i32 @llvm.amdgcn.prng.b32(i32) #0 @@ -29,4 +31,4 @@ define amdgpu_kernel void @prng_b32_constant_100(ptr addrspace(1) %out) #1 { attributes #0 = { nounwind readnone } -attributes #1 = { nounwind }
\ No newline at end of file +attributes #1 = { nounwind } diff --git a/llvm/test/CodeGen/AMDGPU/llvm.exp2.bf16.ll b/llvm/test/CodeGen/AMDGPU/llvm.exp2.bf16.ll index 0f37639..52f6dab 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.exp2.bf16.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.exp2.bf16.ll @@ -16,14 +16,15 @@ define bfloat @v_exp2_bf16(bfloat %in) { ; GFX1200-SDAG-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX1200-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v0 +; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v1 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v1, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0xffffffc0, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v0, v2 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(TRANS32_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v0, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v0, v1 ; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) @@ -90,26 +91,25 @@ define bfloat @v_exp2_fabs_bf16(bfloat %in) { ; GFX1200-SDAG-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX1200-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0x7fff, v0.l -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v0 +; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1200-SDAG-TRUE16-NEXT: v_and_b16 v1.h, 0x7fff, v0.l +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v1 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v1, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0xffffffc0, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v0, v2 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v0, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(TRANS32_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v0, v1 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX1200-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX1200-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -169,26 +169,25 @@ define bfloat @v_exp2_fneg_fabs_bf16(bfloat %in) { ; GFX1200-SDAG-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX1200-SDAG-TRUE16-NEXT: v_or_b16 v0.l, 0x8000, v0.l -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v0 +; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1200-SDAG-TRUE16-NEXT: v_or_b16 v1.h, 0x8000, v0.l +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v1 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v1, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0xffffffc0, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v0, v2 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v0, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(TRANS32_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v0, v1 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX1200-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX1200-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -249,26 +248,25 @@ define bfloat @v_exp2_fneg_bf16(bfloat %in) { ; GFX1200-SDAG-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX1200-SDAG-TRUE16-NEXT: v_xor_b16 v0.l, 0x8000, v0.l -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v0 +; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1200-SDAG-TRUE16-NEXT: v_xor_b16 v1.h, 0x8000, v0.l +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v1 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v1, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0xffffffc0, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v0, v2 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v0, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(TRANS32_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v0, v1 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1 ; GFX1200-SDAG-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_add3_u32 v1, v1, v0, 0x7fff ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h ; GFX1200-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -328,14 +326,15 @@ define bfloat @v_exp2_bf16_fast(bfloat %in) { ; GFX1200-SDAG-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX1200-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v0 +; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v1 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v1, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0xffffffc0, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v0, v2 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(TRANS32_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1) ; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v0, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v0, v1 ; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) @@ -642,47 +641,47 @@ define <2 x bfloat> @v_exp2_fneg_fabs_v2bf16(<2 x bfloat> %in) { ; GFX1200-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 15 -; GFX1200-SDAG-TRUE16-NEXT: v_and_b32_e32 v2, 0x7fff7fff, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX1200-SDAG-TRUE16-NEXT: v_xor_b16 v0.l, 0x8000, v1.l -; GFX1200-SDAG-TRUE16-NEXT: v_xor_b16 v1.l, 0x8000, v2.l -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, 0xc2fc0000, v1 -; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0x42800000, s0 -; GFX1200-SDAG-TRUE16-NEXT: v_dual_add_f32 v1, v1, v2 :: v_dual_lshlrev_b32 v0, 16, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v0 -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0xffffffc0, s0 -; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v1, v1 +; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_xor_b16 v2.h, 0x8000, v1.l +; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v2 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v4, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x42800000, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v3, 0, 0xffffffc0, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v0, v4 -; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v1, v1, v2 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v1, v2, v1 +; GFX1200-SDAG-TRUE16-NEXT: v_xor_b16 v2.h, 0x8000, v0.l ; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v0, v0 -; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v2, v1, 16, 1 -; GFX1200-SDAG-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 -; GFX1200-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(TRANS32_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff -; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v0, v3 +; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v1, v1 +; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, 0xc2fc0000, v2 +; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(TRANS32_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 0x42800000, s0 +; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v1, v1, v3 ; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1 -; GFX1200-SDAG-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX1200-SDAG-TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff -; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1 +; GFX1200-SDAG-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v1 +; GFX1200-SDAG-TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff +; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v2, v0 +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0xffffffc0, s0 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v0, v0 +; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v0, v2 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 +; GFX1200-SDAG-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX1200-SDAG-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX1200-SDAG-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v1, v0 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h +; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v1 ; GFX1200-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX1200-SDAG-FAKE16-LABEL: v_exp2_fneg_fabs_v2bf16: @@ -774,46 +773,44 @@ define <2 x bfloat> @v_exp2_fneg_v2bf16(<2 x bfloat> %in) { ; GFX1200-SDAG-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX1200-SDAG-TRUE16-NEXT: v_xor_b16 v0.l, 0x8000, v0.l -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l -; GFX1200-SDAG-TRUE16-NEXT: v_xor_b16 v0.l, 0x8000, v0.h -; GFX1200-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 +; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1200-SDAG-TRUE16-NEXT: v_xor_b16 v1.h, 0x8000, v0.h +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) ; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v1 ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0xffffffc0, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v3, 0, 0x42800000, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v4, 0, 0x42800000, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX1200-SDAG-TRUE16-NEXT: v_dual_add_f32 v0, v0, v4 :: v_dual_add_f32 v1, v1, v3 +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0x42800000, vcc_lo ; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v3, 0, 0xffffffc0, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v2, v1, v2 +; GFX1200-SDAG-TRUE16-NEXT: v_xor_b16 v1.h, 0x8000, v0.l +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, 0xc2fc0000, v1 +; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 0x42800000, s0 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX1200-SDAG-TRUE16-NEXT: v_add_f32_e32 v0, v1, v0 +; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v1, v2 +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v2, 0, 0xffffffc0, s0 ; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v0, v0 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_2) -; GFX1200-SDAG-TRUE16-NEXT: v_exp_f32_e32 v1, v1 -; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v0, v3 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(TRANS32_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v1, v1, v2 -; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1 -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v2, v1, 16, 1 -; GFX1200-SDAG-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 -; GFX1200-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX1200-SDAG-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v0 -; GFX1200-SDAG-TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff -; GFX1200-SDAG-TRUE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff -; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3) -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(TRANS32_DEP_2) | instskip(NEXT) | instid1(TRANS32_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v1, v1, v3 +; GFX1200-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v0, v2 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1 +; GFX1200-SDAG-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v1 +; GFX1200-SDAG-TRUE16-NEXT: v_bfe_u32 v2, v0, 16, 1 +; GFX1200-SDAG-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v0 ; GFX1200-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0 +; GFX1200-SDAG-TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_add3_u32 v2, v2, v0, 0x7fff ; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo -; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h -; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX1200-SDAG-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v1, v0 +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.l, v0.h +; GFX1200-SDAG-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX1200-SDAG-TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo +; GFX1200-SDAG-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v0, v1 ; GFX1200-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX1200-SDAG-FAKE16-LABEL: v_exp2_fneg_v2bf16: diff --git a/llvm/test/CodeGen/AMDGPU/llvm.log.ll b/llvm/test/CodeGen/AMDGPU/llvm.log.ll index 0d5846a..5634df5 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.log.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.log.ll @@ -6377,28 +6377,99 @@ define float @v_log_f32_from_fpext_bf16(bfloat %src) { ; GFX900-NEXT: v_sub_f32_e32 v0, v0, v1 ; GFX900-NEXT: s_setpc_b64 s[30:31] ; -; GFX1100-LABEL: v_log_f32_from_fpext_bf16: -; GFX1100: ; %bb.0: -; GFX1100-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX1100-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX1100-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 -; GFX1100-NEXT: v_cndmask_b32_e64 v1, 0, 32, vcc_lo -; GFX1100-NEXT: v_ldexp_f32 v0, v0, v1 -; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX1100-NEXT: v_log_f32_e32 v0, v0 -; GFX1100-NEXT: s_waitcnt_depctr 0xfff -; GFX1100-NEXT: v_mul_f32_e32 v1, 0x3f317217, v0 -; GFX1100-NEXT: v_cmp_gt_f32_e64 s0, 0x7f800000, |v0| -; GFX1100-NEXT: v_fma_f32 v2, 0x3f317217, v0, -v1 -; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1100-NEXT: v_fmamk_f32 v2, v0, 0x3377d1cf, v2 -; GFX1100-NEXT: v_add_f32_e32 v1, v1, v2 -; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX1100-NEXT: v_cndmask_b32_e64 v0, v0, v1, s0 -; GFX1100-NEXT: v_cndmask_b32_e64 v1, 0, 0x41b17218, vcc_lo -; GFX1100-NEXT: v_sub_f32_e32 v0, v0, v1 -; GFX1100-NEXT: s_setpc_b64 s[30:31] +; GFX1100-SDAG-TRUE16-LABEL: v_log_f32_from_fpext_bf16: +; GFX1100-SDAG-TRUE16: ; %bb.0: +; GFX1100-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1100-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX1100-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v1 +; GFX1100-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 32, vcc_lo +; GFX1100-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v1, v0 +; GFX1100-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX1100-SDAG-TRUE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-SDAG-TRUE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-SDAG-TRUE16-NEXT: v_mul_f32_e32 v1, 0x3f317217, v0 +; GFX1100-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, 0x7f800000, |v0| +; GFX1100-SDAG-TRUE16-NEXT: v_fma_f32 v2, 0x3f317217, v0, -v1 +; GFX1100-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX1100-SDAG-TRUE16-NEXT: v_fmamk_f32 v2, v0, 0x3377d1cf, v2 +; GFX1100-SDAG-TRUE16-NEXT: v_add_f32_e32 v1, v1, v2 +; GFX1100-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, v0, v1, s0 +; GFX1100-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x41b17218, vcc_lo +; GFX1100-SDAG-TRUE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX1100-SDAG-FAKE16-LABEL: v_log_f32_from_fpext_bf16: +; GFX1100-SDAG-FAKE16: ; %bb.0: +; GFX1100-SDAG-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-SDAG-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX1100-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-SDAG-FAKE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 +; GFX1100-SDAG-FAKE16-NEXT: v_cndmask_b32_e64 v1, 0, 32, vcc_lo +; GFX1100-SDAG-FAKE16-NEXT: v_ldexp_f32 v0, v0, v1 +; GFX1100-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX1100-SDAG-FAKE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-SDAG-FAKE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-SDAG-FAKE16-NEXT: v_mul_f32_e32 v1, 0x3f317217, v0 +; GFX1100-SDAG-FAKE16-NEXT: v_cmp_gt_f32_e64 s0, 0x7f800000, |v0| +; GFX1100-SDAG-FAKE16-NEXT: v_fma_f32 v2, 0x3f317217, v0, -v1 +; GFX1100-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX1100-SDAG-FAKE16-NEXT: v_fmamk_f32 v2, v0, 0x3377d1cf, v2 +; GFX1100-SDAG-FAKE16-NEXT: v_add_f32_e32 v1, v1, v2 +; GFX1100-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-SDAG-FAKE16-NEXT: v_cndmask_b32_e64 v0, v0, v1, s0 +; GFX1100-SDAG-FAKE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x41b17218, vcc_lo +; GFX1100-SDAG-FAKE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-SDAG-FAKE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX1100-GISEL-TRUE16-LABEL: v_log_f32_from_fpext_bf16: +; GFX1100-GISEL-TRUE16: ; %bb.0: +; GFX1100-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-GISEL-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1100-GISEL-TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX1100-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-GISEL-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v1 +; GFX1100-GISEL-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 32, vcc_lo +; GFX1100-GISEL-TRUE16-NEXT: v_ldexp_f32 v0, v1, v0 +; GFX1100-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX1100-GISEL-TRUE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-GISEL-TRUE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-GISEL-TRUE16-NEXT: v_mul_f32_e32 v1, 0x3f317217, v0 +; GFX1100-GISEL-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, 0x7f800000, |v0| +; GFX1100-GISEL-TRUE16-NEXT: v_fma_f32 v2, 0x3f317217, v0, -v1 +; GFX1100-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX1100-GISEL-TRUE16-NEXT: v_fmamk_f32 v2, v0, 0x3377d1cf, v2 +; GFX1100-GISEL-TRUE16-NEXT: v_add_f32_e32 v1, v1, v2 +; GFX1100-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-GISEL-TRUE16-NEXT: v_cndmask_b32_e64 v0, v0, v1, s0 +; GFX1100-GISEL-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x41b17218, vcc_lo +; GFX1100-GISEL-TRUE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX1100-GISEL-FAKE16-LABEL: v_log_f32_from_fpext_bf16: +; GFX1100-GISEL-FAKE16: ; %bb.0: +; GFX1100-GISEL-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-GISEL-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX1100-GISEL-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-GISEL-FAKE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 +; GFX1100-GISEL-FAKE16-NEXT: v_cndmask_b32_e64 v1, 0, 32, vcc_lo +; GFX1100-GISEL-FAKE16-NEXT: v_ldexp_f32 v0, v0, v1 +; GFX1100-GISEL-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX1100-GISEL-FAKE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-GISEL-FAKE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-GISEL-FAKE16-NEXT: v_mul_f32_e32 v1, 0x3f317217, v0 +; GFX1100-GISEL-FAKE16-NEXT: v_cmp_gt_f32_e64 s0, 0x7f800000, |v0| +; GFX1100-GISEL-FAKE16-NEXT: v_fma_f32 v2, 0x3f317217, v0, -v1 +; GFX1100-GISEL-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX1100-GISEL-FAKE16-NEXT: v_fmamk_f32 v2, v0, 0x3377d1cf, v2 +; GFX1100-GISEL-FAKE16-NEXT: v_add_f32_e32 v1, v1, v2 +; GFX1100-GISEL-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-GISEL-FAKE16-NEXT: v_cndmask_b32_e64 v0, v0, v1, s0 +; GFX1100-GISEL-FAKE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x41b17218, vcc_lo +; GFX1100-GISEL-FAKE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-GISEL-FAKE16-NEXT: s_setpc_b64 s[30:31] ; ; R600-LABEL: v_log_f32_from_fpext_bf16: ; R600: ; %bb.0: diff --git a/llvm/test/CodeGen/AMDGPU/llvm.log10.ll b/llvm/test/CodeGen/AMDGPU/llvm.log10.ll index 8006876..8d1a231 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.log10.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.log10.ll @@ -6377,28 +6377,99 @@ define float @v_log10_f32_from_fpext_bf16(bfloat %src) { ; GFX900-NEXT: v_sub_f32_e32 v0, v0, v1 ; GFX900-NEXT: s_setpc_b64 s[30:31] ; -; GFX1100-LABEL: v_log10_f32_from_fpext_bf16: -; GFX1100: ; %bb.0: -; GFX1100-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX1100-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX1100-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 -; GFX1100-NEXT: v_cndmask_b32_e64 v1, 0, 32, vcc_lo -; GFX1100-NEXT: v_ldexp_f32 v0, v0, v1 -; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX1100-NEXT: v_log_f32_e32 v0, v0 -; GFX1100-NEXT: s_waitcnt_depctr 0xfff -; GFX1100-NEXT: v_mul_f32_e32 v1, 0x3e9a209a, v0 -; GFX1100-NEXT: v_cmp_gt_f32_e64 s0, 0x7f800000, |v0| -; GFX1100-NEXT: v_fma_f32 v2, 0x3e9a209a, v0, -v1 -; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX1100-NEXT: v_fmamk_f32 v2, v0, 0x3284fbcf, v2 -; GFX1100-NEXT: v_add_f32_e32 v1, v1, v2 -; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX1100-NEXT: v_cndmask_b32_e64 v0, v0, v1, s0 -; GFX1100-NEXT: v_cndmask_b32_e64 v1, 0, 0x411a209b, vcc_lo -; GFX1100-NEXT: v_sub_f32_e32 v0, v0, v1 -; GFX1100-NEXT: s_setpc_b64 s[30:31] +; GFX1100-SDAG-TRUE16-LABEL: v_log10_f32_from_fpext_bf16: +; GFX1100-SDAG-TRUE16: ; %bb.0: +; GFX1100-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1100-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX1100-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v1 +; GFX1100-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 32, vcc_lo +; GFX1100-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v1, v0 +; GFX1100-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX1100-SDAG-TRUE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-SDAG-TRUE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-SDAG-TRUE16-NEXT: v_mul_f32_e32 v1, 0x3e9a209a, v0 +; GFX1100-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, 0x7f800000, |v0| +; GFX1100-SDAG-TRUE16-NEXT: v_fma_f32 v2, 0x3e9a209a, v0, -v1 +; GFX1100-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX1100-SDAG-TRUE16-NEXT: v_fmamk_f32 v2, v0, 0x3284fbcf, v2 +; GFX1100-SDAG-TRUE16-NEXT: v_add_f32_e32 v1, v1, v2 +; GFX1100-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, v0, v1, s0 +; GFX1100-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x411a209b, vcc_lo +; GFX1100-SDAG-TRUE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX1100-SDAG-FAKE16-LABEL: v_log10_f32_from_fpext_bf16: +; GFX1100-SDAG-FAKE16: ; %bb.0: +; GFX1100-SDAG-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-SDAG-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX1100-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-SDAG-FAKE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 +; GFX1100-SDAG-FAKE16-NEXT: v_cndmask_b32_e64 v1, 0, 32, vcc_lo +; GFX1100-SDAG-FAKE16-NEXT: v_ldexp_f32 v0, v0, v1 +; GFX1100-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX1100-SDAG-FAKE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-SDAG-FAKE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-SDAG-FAKE16-NEXT: v_mul_f32_e32 v1, 0x3e9a209a, v0 +; GFX1100-SDAG-FAKE16-NEXT: v_cmp_gt_f32_e64 s0, 0x7f800000, |v0| +; GFX1100-SDAG-FAKE16-NEXT: v_fma_f32 v2, 0x3e9a209a, v0, -v1 +; GFX1100-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX1100-SDAG-FAKE16-NEXT: v_fmamk_f32 v2, v0, 0x3284fbcf, v2 +; GFX1100-SDAG-FAKE16-NEXT: v_add_f32_e32 v1, v1, v2 +; GFX1100-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-SDAG-FAKE16-NEXT: v_cndmask_b32_e64 v0, v0, v1, s0 +; GFX1100-SDAG-FAKE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x411a209b, vcc_lo +; GFX1100-SDAG-FAKE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-SDAG-FAKE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX1100-GISEL-TRUE16-LABEL: v_log10_f32_from_fpext_bf16: +; GFX1100-GISEL-TRUE16: ; %bb.0: +; GFX1100-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-GISEL-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1100-GISEL-TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX1100-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-GISEL-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v1 +; GFX1100-GISEL-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 32, vcc_lo +; GFX1100-GISEL-TRUE16-NEXT: v_ldexp_f32 v0, v1, v0 +; GFX1100-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX1100-GISEL-TRUE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-GISEL-TRUE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-GISEL-TRUE16-NEXT: v_mul_f32_e32 v1, 0x3e9a209a, v0 +; GFX1100-GISEL-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, 0x7f800000, |v0| +; GFX1100-GISEL-TRUE16-NEXT: v_fma_f32 v2, 0x3e9a209a, v0, -v1 +; GFX1100-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX1100-GISEL-TRUE16-NEXT: v_fmamk_f32 v2, v0, 0x3284fbcf, v2 +; GFX1100-GISEL-TRUE16-NEXT: v_add_f32_e32 v1, v1, v2 +; GFX1100-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-GISEL-TRUE16-NEXT: v_cndmask_b32_e64 v0, v0, v1, s0 +; GFX1100-GISEL-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x411a209b, vcc_lo +; GFX1100-GISEL-TRUE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX1100-GISEL-FAKE16-LABEL: v_log10_f32_from_fpext_bf16: +; GFX1100-GISEL-FAKE16: ; %bb.0: +; GFX1100-GISEL-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-GISEL-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX1100-GISEL-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-GISEL-FAKE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 +; GFX1100-GISEL-FAKE16-NEXT: v_cndmask_b32_e64 v1, 0, 32, vcc_lo +; GFX1100-GISEL-FAKE16-NEXT: v_ldexp_f32 v0, v0, v1 +; GFX1100-GISEL-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX1100-GISEL-FAKE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-GISEL-FAKE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-GISEL-FAKE16-NEXT: v_mul_f32_e32 v1, 0x3e9a209a, v0 +; GFX1100-GISEL-FAKE16-NEXT: v_cmp_gt_f32_e64 s0, 0x7f800000, |v0| +; GFX1100-GISEL-FAKE16-NEXT: v_fma_f32 v2, 0x3e9a209a, v0, -v1 +; GFX1100-GISEL-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX1100-GISEL-FAKE16-NEXT: v_fmamk_f32 v2, v0, 0x3284fbcf, v2 +; GFX1100-GISEL-FAKE16-NEXT: v_add_f32_e32 v1, v1, v2 +; GFX1100-GISEL-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-GISEL-FAKE16-NEXT: v_cndmask_b32_e64 v0, v0, v1, s0 +; GFX1100-GISEL-FAKE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x411a209b, vcc_lo +; GFX1100-GISEL-FAKE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-GISEL-FAKE16-NEXT: s_setpc_b64 s[30:31] ; ; R600-LABEL: v_log10_f32_from_fpext_bf16: ; R600: ; %bb.0: diff --git a/llvm/test/CodeGen/AMDGPU/llvm.log2.ll b/llvm/test/CodeGen/AMDGPU/llvm.log2.ll index c1ac74e..7ca72bf 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.log2.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.log2.ll @@ -3830,20 +3830,67 @@ define float @v_log2_f32_from_fpext_bf16(bfloat %src) { ; GFX900-NEXT: v_sub_f32_e32 v0, v0, v1 ; GFX900-NEXT: s_setpc_b64 s[30:31] ; -; GFX1100-LABEL: v_log2_f32_from_fpext_bf16: -; GFX1100: ; %bb.0: -; GFX1100-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX1100-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX1100-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 -; GFX1100-NEXT: v_cndmask_b32_e64 v2, 0, 32, vcc_lo -; GFX1100-NEXT: v_cndmask_b32_e64 v1, 0, 0x42000000, vcc_lo -; GFX1100-NEXT: v_ldexp_f32 v0, v0, v2 -; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX1100-NEXT: v_log_f32_e32 v0, v0 -; GFX1100-NEXT: s_waitcnt_depctr 0xfff -; GFX1100-NEXT: v_sub_f32_e32 v0, v0, v1 -; GFX1100-NEXT: s_setpc_b64 s[30:31] +; GFX1100-SDAG-TRUE16-LABEL: v_log2_f32_from_fpext_bf16: +; GFX1100-SDAG-TRUE16: ; %bb.0: +; GFX1100-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1100-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX1100-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-SDAG-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v1 +; GFX1100-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 32, vcc_lo +; GFX1100-SDAG-TRUE16-NEXT: v_ldexp_f32 v0, v1, v0 +; GFX1100-SDAG-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x42000000, vcc_lo +; GFX1100-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX1100-SDAG-TRUE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-SDAG-TRUE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-SDAG-TRUE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX1100-SDAG-FAKE16-LABEL: v_log2_f32_from_fpext_bf16: +; GFX1100-SDAG-FAKE16: ; %bb.0: +; GFX1100-SDAG-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-SDAG-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX1100-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX1100-SDAG-FAKE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 +; GFX1100-SDAG-FAKE16-NEXT: v_cndmask_b32_e64 v2, 0, 32, vcc_lo +; GFX1100-SDAG-FAKE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x42000000, vcc_lo +; GFX1100-SDAG-FAKE16-NEXT: v_ldexp_f32 v0, v0, v2 +; GFX1100-SDAG-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX1100-SDAG-FAKE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-SDAG-FAKE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-SDAG-FAKE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-SDAG-FAKE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX1100-GISEL-TRUE16-LABEL: v_log2_f32_from_fpext_bf16: +; GFX1100-GISEL-TRUE16: ; %bb.0: +; GFX1100-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-GISEL-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 +; GFX1100-GISEL-TRUE16-NEXT: v_mov_b16_e32 v1.h, v0.l +; GFX1100-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX1100-GISEL-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v1 +; GFX1100-GISEL-TRUE16-NEXT: v_cndmask_b32_e64 v0, 0, 32, vcc_lo +; GFX1100-GISEL-TRUE16-NEXT: v_ldexp_f32 v0, v1, v0 +; GFX1100-GISEL-TRUE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x42000000, vcc_lo +; GFX1100-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX1100-GISEL-TRUE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-GISEL-TRUE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-GISEL-TRUE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; GFX1100-GISEL-FAKE16-LABEL: v_log2_f32_from_fpext_bf16: +; GFX1100-GISEL-FAKE16: ; %bb.0: +; GFX1100-GISEL-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GFX1100-GISEL-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; GFX1100-GISEL-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX1100-GISEL-FAKE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0 +; GFX1100-GISEL-FAKE16-NEXT: v_cndmask_b32_e64 v2, 0, 32, vcc_lo +; GFX1100-GISEL-FAKE16-NEXT: v_cndmask_b32_e64 v1, 0, 0x42000000, vcc_lo +; GFX1100-GISEL-FAKE16-NEXT: v_ldexp_f32 v0, v0, v2 +; GFX1100-GISEL-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX1100-GISEL-FAKE16-NEXT: v_log_f32_e32 v0, v0 +; GFX1100-GISEL-FAKE16-NEXT: s_waitcnt_depctr 0xfff +; GFX1100-GISEL-FAKE16-NEXT: v_sub_f32_e32 v0, v0, v1 +; GFX1100-GISEL-FAKE16-NEXT: s_setpc_b64 s[30:31] ; ; R600-LABEL: v_log2_f32_from_fpext_bf16: ; R600: ; %bb.0: diff --git a/llvm/test/CodeGen/AMDGPU/mad-mix-hi.ll b/llvm/test/CodeGen/AMDGPU/mad-mix-hi.ll index 6e94896..c0fb145 100644 --- a/llvm/test/CodeGen/AMDGPU/mad-mix-hi.ll +++ b/llvm/test/CodeGen/AMDGPU/mad-mix-hi.ll @@ -193,13 +193,22 @@ define <2 x half> @v_mad_mixhi_f16_f16lo_f16lo_f16lo_reglo(half %src0, half %src } define i32 @v_mad_mixhi_f16_f16lo_f16lo_f16lo_intpack(half %src0, half %src1, half %src2) #0 { -; SDAG-GFX11-LABEL: v_mad_mixhi_f16_f16lo_f16lo_f16lo_intpack: -; SDAG-GFX11: ; %bb.0: -; SDAG-GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; SDAG-GFX11-NEXT: v_fma_mixlo_f16 v0, v0, v1, v2 op_sel_hi:[1,1,1] -; SDAG-GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; SDAG-GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; SDAG-GFX11-NEXT: s_setpc_b64 s[30:31] +; SDAG-GFX11-TRUE16-LABEL: v_mad_mixhi_f16_f16lo_f16lo_f16lo_intpack: +; SDAG-GFX11-TRUE16: ; %bb.0: +; SDAG-GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; SDAG-GFX11-TRUE16-NEXT: v_fma_mixlo_f16 v1, v0, v1, v2 op_sel_hi:[1,1,1] +; SDAG-GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; SDAG-GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; SDAG-GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l +; SDAG-GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; SDAG-GFX11-FAKE16-LABEL: v_mad_mixhi_f16_f16lo_f16lo_f16lo_intpack: +; SDAG-GFX11-FAKE16: ; %bb.0: +; SDAG-GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; SDAG-GFX11-FAKE16-NEXT: v_fma_mixlo_f16 v0, v0, v1, v2 op_sel_hi:[1,1,1] +; SDAG-GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; SDAG-GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; SDAG-GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31] ; ; SDAG-GFX9-LABEL: v_mad_mixhi_f16_f16lo_f16lo_f16lo_intpack: ; SDAG-GFX9: ; %bb.0: @@ -265,13 +274,22 @@ define i32 @v_mad_mixhi_f16_f16lo_f16lo_f16lo_intpack(half %src0, half %src1, ha } define i32 @v_mad_mixhi_f16_f16lo_f16lo_f16lo_intpack_sext(half %src0, half %src1, half %src2) #0 { -; SDAG-GFX11-LABEL: v_mad_mixhi_f16_f16lo_f16lo_f16lo_intpack_sext: -; SDAG-GFX11: ; %bb.0: -; SDAG-GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; SDAG-GFX11-NEXT: v_fma_mixlo_f16 v0, v0, v1, v2 op_sel_hi:[1,1,1] -; SDAG-GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) -; SDAG-GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; SDAG-GFX11-NEXT: s_setpc_b64 s[30:31] +; SDAG-GFX11-TRUE16-LABEL: v_mad_mixhi_f16_f16lo_f16lo_f16lo_intpack_sext: +; SDAG-GFX11-TRUE16: ; %bb.0: +; SDAG-GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; SDAG-GFX11-TRUE16-NEXT: v_fma_mixlo_f16 v1, v0, v1, v2 op_sel_hi:[1,1,1] +; SDAG-GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0 +; SDAG-GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; SDAG-GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l +; SDAG-GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] +; +; SDAG-GFX11-FAKE16-LABEL: v_mad_mixhi_f16_f16lo_f16lo_f16lo_intpack_sext: +; SDAG-GFX11-FAKE16: ; %bb.0: +; SDAG-GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; SDAG-GFX11-FAKE16-NEXT: v_fma_mixlo_f16 v0, v0, v1, v2 op_sel_hi:[1,1,1] +; SDAG-GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; SDAG-GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0 +; SDAG-GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31] ; ; SDAG-GFX9-LABEL: v_mad_mixhi_f16_f16lo_f16lo_f16lo_intpack_sext: ; SDAG-GFX9: ; %bb.0: @@ -569,3 +587,4 @@ attributes #0 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" attributes #1 = { nounwind readnone speculatable } ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line: ; GISEL-GFX11-FAKE16: {{.*}} +; SDAG-GFX11: {{.*}} diff --git a/llvm/test/CodeGen/AMDGPU/maximumnum.bf16.ll b/llvm/test/CodeGen/AMDGPU/maximumnum.bf16.ll index 6246f2f..ca16e25 100644 --- a/llvm/test/CodeGen/AMDGPU/maximumnum.bf16.ll +++ b/llvm/test/CodeGen/AMDGPU/maximumnum.bf16.ll @@ -118,34 +118,29 @@ define bfloat @v_maximumnum_bf16(bfloat %x, bfloat %y) { ; GFX11-TRUE16-LABEL: v_maximumnum_bf16: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v1, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_bf16: @@ -181,40 +176,34 @@ define bfloat @v_maximumnum_bf16(bfloat %x, bfloat %y) { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v2, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v1, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_bf16: @@ -339,21 +328,21 @@ define bfloat @v_maximumnum_bf16_nnan(bfloat %x, bfloat %y) { ; GFX11-TRUE16-LABEL: v_maximumnum_bf16_nnan: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v3, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v3, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.l, v0.l, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v0.l, s0 ; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_bf16_nnan: @@ -381,25 +370,25 @@ define bfloat @v_maximumnum_bf16_nnan(bfloat %x, bfloat %y) { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v3, v2 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v3, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.l, v0.l, vcc_lo ; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v0.l, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_bf16_nnan: @@ -630,58 +619,46 @@ define <2 x bfloat> @v_maximumnum_v2bf16(<2 x bfloat> %x, <2 x bfloat> %y) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v5, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v4.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v4, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.h, v4.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v4.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v1.l, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.l, v3.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v3, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v3.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v3.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v2.l, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v4, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v1.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v2bf16: @@ -738,62 +715,56 @@ define <2 x bfloat> @v_maximumnum_v2bf16(<2 x bfloat> %x, <2 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v5, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v1.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v4, v6 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v4.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v4, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.h, v4.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v1.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v1.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.l, v3.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v3, v4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v3.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v3.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v2.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v2bf16: @@ -1012,34 +983,29 @@ define <2 x bfloat> @v_maximumnum_v2bf16_nnan(<2 x bfloat> %x, <2 x bfloat> %y) ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v0 ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.h -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.h ; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v0.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, s2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v3.h, v0.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v1.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.h, v0.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.h, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.h, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v2bf16_nnan: @@ -1085,36 +1051,35 @@ define <2 x bfloat> @v_maximumnum_v2bf16_nnan(<2 x bfloat> %x, <2 x bfloat> %y) ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l ; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v0 ; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.h -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.h ; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v0.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v3.h, v0.l, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v1.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.h, v0.l, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, s0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.h, v0.h, s1 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.h, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v2bf16_nnan: @@ -1444,66 +1409,67 @@ define <3 x bfloat> @v_maximumnum_v3bf16(<3 x bfloat> %x, <3 x bfloat> %y) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v6, v6 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.h, v4.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v8 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v7, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v9, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v2.l, v0.l, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v4.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v7, v7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v7, v7 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v5.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v5.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v5.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v3, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.h, v4.h, s2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v3bf16: @@ -1575,77 +1541,80 @@ define <3 x bfloat> @v_maximumnum_v3bf16(<3 x bfloat> %x, <3 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v6, v6 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.h, v4.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v7, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v8 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v7, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v9, v11 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v7, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v2.l, v0.l, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v4.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v5.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v5.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v5.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v3, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v0.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v0, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v4.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.h, v4.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v3bf16: @@ -1939,41 +1908,40 @@ define <3 x bfloat> @v_maximumnum_v3bf16_nnan(<3 x bfloat> %x, <3 x bfloat> %y) ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v3 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v4 ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v7, v6 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v3.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v9, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.h -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v0.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v3.l, v1.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.h, v1.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v5 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.h, v1.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.l, v0.l, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v8 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.h, v0.l, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v2.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v0.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v0.h, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v3bf16_nnan: @@ -2029,48 +1997,50 @@ define <3 x bfloat> @v_maximumnum_v3bf16_nnan(<3 x bfloat> %x, <3 x bfloat> %y) ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v3 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v4 ; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v7, v6 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v3.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v9, v8 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v3.l, v1.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.l, v0.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.h -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v0.h, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.h, v1.l, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v8 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v2.h -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.h, v1.l, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.h, v0.l, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v2.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v8 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v0.l, s0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v0.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v0.h, s0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v3bf16_nnan: @@ -2507,85 +2477,83 @@ define <4 x bfloat> @v_maximumnum_v4bf16(<4 x bfloat> %x, <4 x bfloat> %y) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v0.h, v2.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v4.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v10, v10 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v1.h, v3.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v10, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v5.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v9, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v7.l, v6.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v8.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v11, v12 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v9, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v7.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v0.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v8, v8 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v2.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v7.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v5.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v1.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v8.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v8, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v8.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v8.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v3, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v7.h, s4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v5.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v5.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v4bf16: @@ -2680,99 +2648,98 @@ define <4 x bfloat> @v_maximumnum_v4bf16(<4 x bfloat> %x, <4 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v0.h, v2.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v4.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v10, v10 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v10, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v1.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v5.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v6.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v9, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v1.h, v3.h, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v7.l, v6.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v8.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v11, v12 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v7.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v9, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v1.l, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v8, v8 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v1.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v7.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v2.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v7.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v5.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v0.h, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v1.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v8.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v8.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v8, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v8.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v8.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v3, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v7.h, s4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v0.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v0, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v6.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v5.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v5.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v0.h, s1 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v4bf16: @@ -3158,53 +3125,52 @@ define <4 x bfloat> @v_maximumnum_v4bf16_nnan(<4 x bfloat> %x, <4 x bfloat> %y) ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v3 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v2 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v5, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v7, v6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v1.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.l, v1.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v1.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v7, v6 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v1.h, s4 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v6, v8 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v10, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, s4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.h, v1.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.h, v1.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v2.h, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.h -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v3.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v3.h -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v1.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v3.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.h, v1.h, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v5.h, v0.l, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v2.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.h, v0.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v2.h, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v3.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v8 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.h, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v1.h, s2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.h, v0.h, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v4bf16_nnan: @@ -3273,62 +3239,63 @@ define <4 x bfloat> @v_maximumnum_v4bf16_nnan(<4 x bfloat> %x, <4 x bfloat> %y) ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v3 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.l ; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v2 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v5, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v7, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v1.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.l, v1.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v1.h, s4 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v6, v8 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.h -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v1.l, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v10, v7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v7, v6 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, s4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.h, v1.l, s0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.h, v1.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v2.h, v0.h, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.h -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v3.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v3.h -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v1.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v3.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v4 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.h, v1.h, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.h -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v3.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v5.h, v0.l, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v0.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v2.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v8 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.h, v0.l, s0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v2.h, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.h, v0.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v1.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.h, v0.h, s0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v4bf16_nnan: @@ -3957,125 +3924,120 @@ define <6 x bfloat> @v_maximumnum_v6bf16(<6 x bfloat> %x, <6 x bfloat> %y) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v2 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v0 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v7, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v9, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v2.h, v5.h, vcc_lo ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v5 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v10, v10 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v5.h, v6.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v1.h, v4.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v5.h, v9.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v9.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v13, v13 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v9, v8 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v13, v13 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v8.h, v9.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v9.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v6.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v9 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v1.h, v4.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v4.h, v9.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v9.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v9, v10 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v10.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v10.h, v9.h, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v9.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v7.l ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v4.h, v8.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v9.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v5.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v0.h, v3.h, s2 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v13, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v14 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v8.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v6.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v9.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v7.l, v6.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v12, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v3.h, v10.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v10.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v11.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v9.l, v8.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v12.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v11.l, v6.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v13.l, v8.l, s3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.h, v7.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v9.l, s4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v2.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v8, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v15 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v14, v16 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v11.l, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v8, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v12.l, v10.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v10.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v1.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v11, v7 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v0.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v5.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v12.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v10, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v14, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v4.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v3.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v5.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v11 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 ; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v13.l, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v7.l, v3.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v0.h, v3.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v3.h, v9.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v9.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v9, v11 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v11.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v11.h, v9.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v8.l, v9.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v8.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v2.l, v5.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v5.l, v12.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v12, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v9.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v9.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v12.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.h, v9.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v4.l, s3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v9.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v12, v12 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v4.l, v1.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v1.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v2.h, v11.h, s7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v4.h, s0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v1, v9 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v9.h, v1.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v9.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v1.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v5, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.h, v10.h, s6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.h, v9.h, s3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v3.l, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v9.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.l, v1.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v3.l, v0.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v0, v9 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v9.h, v0.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v9.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v8.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v0.h, v9.h, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.h, v8.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v6.l, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s4 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v6bf16: @@ -4206,142 +4168,141 @@ define <6 x bfloat> @v_maximumnum_v6bf16(<6 x bfloat> %x, <6 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v2 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v5 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v1 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v4 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v0 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v7, v7 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v9, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v9.l +; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v2.h, v5.h, vcc_lo ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v5 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v10, v10 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v5.h, v6.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v5.h, v9.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v9.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v13, v13 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v9, v8 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v13, v13 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v1.h, v4.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v7.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v4.h, v8.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v9.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v5.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v0.h, v3.h, s2 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v13, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v14 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v8.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v6.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v9.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v8.h, v9.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v7.l, v6.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v12, v13 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v9.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v6.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v9 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v3.h, v10.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v10.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v7.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v11.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v1.h, v4.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v9.l, v8.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, v12.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v11.l, v6.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v15 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v13.l, v8.l, s3 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v13.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.h, v7.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v9.l, s4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v8, v8 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v15 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v14, v16 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v4.h, v9.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v9.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v9, v10 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v10.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v11.l, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v8, v8 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v12.l, v10.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v10.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v10.h, v9.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v9.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v10.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v1.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v11, v7 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v0.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v12.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v9.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v7.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v9 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v0.h, v3.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v5.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v12.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v10, v9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v3.h, v9.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v9.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v9, v11 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v11.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v11.h, v9.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v14, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v8 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v4.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v1.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v8.l, v9.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v8.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v2.l, v5.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v5.l, v12.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v12.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v12, v9 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v9.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v9.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v12.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v5, v5 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.h, v9.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v4.l, s3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v9.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v12, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v4.l, v1.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v1.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v2.h, v11.h, s7 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v1, v9 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v9.h, v1.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v9.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v1.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v5, v5 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.h, v10.h, s6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.h, v9.h, s3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v3.l, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v9.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.l, v1.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v3.l, v0.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v0.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v3.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.h, s3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v0, v9 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v5.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.l, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v9.h, v0.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v9.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v8.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v11 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v9 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v0.h, v9.h, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.h, v8.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v6.l, v0.h, s1 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v13.l, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v7.l, v3.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s4 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v6bf16: @@ -5219,171 +5180,160 @@ define <8 x bfloat> @v_maximumnum_v8bf16(<8 x bfloat> %x, <8 x bfloat> %y) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v6 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff0000, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v3.h, v7.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v2.h, v6.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v3.h, v7.h, vcc_lo +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v7.h, v8.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v6.h, v9.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v1.h, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v8.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v9.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v15, v17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v5.h, v12.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v13, v18 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v10.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v10.l, v8.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v11.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v11.l, v9.l, s3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v14.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v13.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v13.l, v8.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v15.l, v9.l, s4 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v7.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v17, v17 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v11 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v17, v17 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v11.h, v12.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.l, v12.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v8.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v2.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v6.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0, v13.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v13.h, v12.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.l, v12.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v9.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v14, v14 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v1.h, v5.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v5.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v14 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v14.h, v12.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.l, v12.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v10.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v16, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v3 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v0.h, v4.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v4.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v15 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v15.h, v12.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v12.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v11.l, v12.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v11.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v3.l, v7.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v7.l, v16.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v16.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v16, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v12.h, v16.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.l, v16.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v7, v7 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v1 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v3.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.l, v6.l, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v12.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v16, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v6.l, v2.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v2.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v10.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v16, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v6.h, s0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v2, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v12.h, v2.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v12.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v2.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v7, v7 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.h, v13.h, s8 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.h, v12.h, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v5.l, s5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v12.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.l, v2.h, s2 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v19 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v8.h, v11.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v20 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v0.h, v4.h, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v16, v16 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v11, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v13.l, v8.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v15.l, v9.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.h, v10.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v14.l, v12.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v7 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v7.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v9.l, v12.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v14.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v16 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v13, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v14.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v3.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v8.l, v10.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v6.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v11.l, v10.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v12, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.h, v8.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v7.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v5 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v5.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.l, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v10 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v15, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v3.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v4.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v1.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v6.l, v2.l, s3 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.h, v7.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v14, v13 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v5.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v6.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v7.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v13 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v11.l, v3.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v4.l, v4.h, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v1.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v7.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v5.l, v1.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v5.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s5, v1, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v12.h, v1.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v12.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v1.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v7, v7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v12.h, s5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v4.l, s6 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v12.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v9 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v8 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v4.l, v0.h, s7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s6, v0, v12 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0, v12.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v12.h, v0.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v11.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.h, v11.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v14.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v15.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v12.h, s9 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v10.h, v14.h, s6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v15.h, s7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v8.l, v1.h, s1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v10.l, v0.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v11.l, v5.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v4.l, s1 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v8bf16: @@ -5546,201 +5496,187 @@ define <8 x bfloat> @v_maximumnum_v8bf16(<8 x bfloat> %x, <8 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v7 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v6 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v6 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff0000, v5 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v12.l +; GFX12-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v3.h, v7.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v2.h, v6.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v3.h, v7.h, vcc_lo +; GFX12-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v4 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v7.h, v8.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v8.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v6.h, v9.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v10.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v1.h, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v8.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v9.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v15, v17 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v5.h, v12.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v13, v18 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v10.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v10.l, v8.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v11.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v11.l, v9.l, s3 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v20.l, v14.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v13.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v13.l, v8.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v15.l, v9.l, s4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v10.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v19 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v8.h, v11.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v20 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v0.h, v4.h, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v16, v16 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v11, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v7.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v17, v17 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v6 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v11 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v17, v17 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v13.l, v8.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v11.h, v12.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v15.l, v9.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.h, v10.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v14.l, v12.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v7 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v12.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v8.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v7.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, v9.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.l, v12.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v8.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v12 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v9.l, v12.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v14.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v2.h, v6.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v16 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v13, v15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v6.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0, v13.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v14.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v3.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v11 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v8.l, v10.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v10.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v6.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v14 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v11.l, v10.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v8.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v12, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.h, v8.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v11.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v6.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v7.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v5 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v13.h, v12.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v5.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.l, v12.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v9.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v14, v14 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v12 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v1.h, v5.h, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v5.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v14 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v14.h, v12.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.l, v12.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v10.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v16, v16 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v3 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v12 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.l, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v0.h, v4.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v10 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v15, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v4.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v15 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v3.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v4.l, v0.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v1.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v7.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v15.h, v12.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, v12.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v6.l, v2.l, s3 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v11.l, v12.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v11.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.h, v7.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v8.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v3.l, v7.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v7.l, v16.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v16.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v16, v12 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v12.h, v16.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.l, v16.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v7, v7 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v1 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v5 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v3.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.l, v6.l, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v12.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v16, v16 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v12 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v6.l, v2.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v2.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v16, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v2, v12 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v12.h, v2.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v12.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v2.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v7, v7 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.h, v13.h, s8 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v14, v13 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v6.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v5.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.h, v12.h, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v5.l, s5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v12.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.l, v2.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v12 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v5.l, v1.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v1.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v6.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v2.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v5.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v5.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s5, v1, v12 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v12.h, v1.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v12.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v1.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v7, v7 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v12.h, s5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v4.l, s6 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v12.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v12 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v4.l, v0.h, s7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v0.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v7.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v5.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s6, v0, v12 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0, v12.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v13 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v11.l, v3.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v12.h, v0.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v11.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v4.l, v4.h, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v1.h, s3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.h, v11.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v14.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v15.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v12.h, s9 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v0.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v10.h, v14.h, s6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v15.h, s7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v8.l, v1.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v10.l, v0.h, s3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v7.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v9 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v8 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v11.l, v5.l, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v4.l, s1 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v8bf16: @@ -7352,341 +7288,314 @@ define <16 x bfloat> @v_maximumnum_v16bf16(<16 x bfloat> %x, <16 x bfloat> %y) { ; GFX11-TRUE16-LABEL: v_maximumnum_v16bf16: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, v7 :: v_dual_mov_b32 v17, v6 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v18, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v7 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff0000, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v6 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v14 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v17 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v13 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v15.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v15.h, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v19 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v17.h, v14.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v18 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v14.h, v7.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v22, v21 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v7.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v19.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v18.h, v13.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.l, v6.l, v5.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v22.l, v13.h, v20.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v21.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v21.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v24, v23 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v22.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v20.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v6.l, s1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v23.l, v19.l, v7.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v26 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v19.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v23.l, v7.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v21.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v12 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v23.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v4.h, v12.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v22.l, v20.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v20.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v19.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v12.h, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v21 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v6.l, v20.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v11 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v22.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v3.h, v11.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v23.l, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v26, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v19.l, v22.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.l, v11.h, v20.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v20.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v7.l, v5.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v21.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v19.l, v5.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v13 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v12 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v20, v20 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v7.h, v15.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v21, v21 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v16.l +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.h, v15.h, v16.h, vcc_lo +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v22, v22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v3 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v16, v17 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v16.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v19.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v7.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v22, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v10.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v16.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.h, v16.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v23, v23 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v24, v24 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v22.l, v21.l, v20.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v20.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v7.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v10.h, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v23 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v22.l, v20.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v9 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v23.l, v1.h, v9.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.h, v19.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v9.h, v23.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v20.l, v21.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v27, v25 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v5.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.h, v22.l, v19.l, vcc_lo -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v7.l, v6.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v20.l, v6.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v20.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v21, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v0.h, v8.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v7.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.l, v5.l, v23.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v23.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v8.h, v19.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v16 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v24 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v21.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v21.l, v23.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v22, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.l, v16.l, v15.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v22, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v20.l, v6.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v23, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v15.l, v16.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v7.l, v19.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v16.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v22, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v14 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.l, v14.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v22, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v19.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v25 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v24, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v17.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v21.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v7.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v22 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v6.l, v16.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v16.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.l, v16.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v6.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v22, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v18.l, v13.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v6.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.l, v14.l, v17.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v17.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v13.l, v15.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v24 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v16.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v16.l, v17.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v14.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v12 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v12.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.h, v14.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v18, v17 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v6.l, v15.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v11.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v5.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v15.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v18, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v3.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v21 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v5.l, v15.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v17, v17 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v6.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v17.l, v16.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v17.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff0000, v9 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v6.h, v14.h, s0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v8 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v29, v29 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v15 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v18.h, v14.h, v16.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v16.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v7 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v32, v32 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v16, v18 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v14 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v18.l, v18.h, v16.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v32, v32 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.h, v18.l, v16.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v18.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v20, v20 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v32, v32 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v5.h, v13.h, s1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v32, v32 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v11 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v25.h, v13.h, v16.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v16.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v32, v32 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v10 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v16, v25 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0, v25.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v32, v32 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v25.h, v16.h, s1 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.h, v20.l, v16.h, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v20.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v21, v21 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v32, v32 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v16.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v10.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v4.h, v12.h, s2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v26.h, v12.h, v16.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v16.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v16, v26 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.l, v26.h, v16.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.h, v21.l, v16.h, s3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v21.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v22, v22 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v3.h, v11.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v27.h, v11.h, v16.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v16.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v16, v27 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v22.l, v27.h, v16.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v22.h, v22.l, v16.h, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v22.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v23, v23 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v2.h, v10.h, s4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v28.h, v10.h, v16.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v16.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v16, v28 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v23.l, v28.h, v16.h, s4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v23.h, v23.l, v16.h, s5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v23.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v24, v24 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v1.h, v9.h, s5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v29.h, v9.h, v16.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v16.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s5, v16, v29 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v24.l, v29.h, v16.h, s5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v24.h, v24.l, v16.h, s6 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v24.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v30, v30 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v0.h, v8.h, s6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.h, v8.h, v16.h, s7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v16.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s6, v16, v30 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v30.h, v16.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v25.l, v15.h, v16.h, s7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v15.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v15.l, s6 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v16.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v7.h, s8 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0, v7.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s7, v7, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v16.h, v7.h, s7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v7.l, v7.h, s8 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v6.h, v16.h, s7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v7.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v14.l, s8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v14.l, v6.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v14.h, s7 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s8, v6, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v16.h, v6.h, s8 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v6.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v5.h, v16.h, s8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v6.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.l, v13.l, s9 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v13.l, v5.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0, v5.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v13.h, s8 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s9, v5, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v5.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v5.l, v5.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v4.h, v16.h, s9 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v5.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v12.l, s10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v12.l, v4.h, s11 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v12.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s10, v4, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v16.h, v4.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v4.l, v4.h, s11 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v3.h, v16.h, s10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.l, v11.l, s11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v11.l, v3.h, s12 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v11.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s11, v3, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v16.h, v3.h, s11 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.l, v3.h, s12 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v2.h, v16.h, s11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v3.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.l, v10.l, s12 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v10.l, v2.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v10.h, s11 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s12, v2, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v16.h, v2.h, s12 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v2.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v1.h, v16.h, s12 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v9.l, s13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v9.l, v1.h, s14 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v31, v31 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0, v1.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v9.h, s12 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s13, v1, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.h, v0.l, v8.l, s14 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0, v16.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v31, v31 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v16.h, v1.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0, v17.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.l, v1.h, s15 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v19.l, v17.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0, v18.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v20.h, v25.h, s17 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0, v32.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v16.h, s16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v17.l, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v19.h, v18.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v26.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v20.l, v1.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v8.l, v32.h, s14 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v18.l, v0.h, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v21.h, v26.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v28.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v27.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v32, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s13 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v21.l, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v29.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v23.h, v28.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v16.h, v32.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v22.h, v27.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v24.h, v29.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v30.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v8.l, v32.h, s15 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v23.l, v1.h, s4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v22.l, v0.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v24.l, v8.h, s5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v25.l, v30.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v9.l, v16.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v8.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v15.h, v0.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v16 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v17, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v12.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v16.l, v4.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v17, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v13.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v10.l, v2.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v11.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v21, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v9.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v8.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v21, v21 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v10.l, v3.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v8.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v12.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v22, v17 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v0.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v11.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.l, v2.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v11.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v18, v17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v22, v21 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v9.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v13.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v3.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v4.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v8.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v9.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v11.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v9.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v21 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v17 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.l, v11.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v2.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v8.l, v1.h, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v10.l, v0.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v13.l, v2.h, s3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, v16 :: v_dual_mov_b32 v1, v14 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, v15 :: v_dual_mov_b32 v3, v20 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v19 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v8.h, vcc_lo ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v16bf16: @@ -8005,406 +7914,355 @@ define <16 x bfloat> @v_maximumnum_v16bf16(<16 x bfloat> %x, <16 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v16, v7 :: v_dual_mov_b32 v17, v6 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v18, v5 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v15 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v7 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff0000, v15 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v6 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v14 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v16 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v17 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v13 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v15.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v19.l, v5.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v15.h, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v19 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v6.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v17.h, v14.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v14.h, v7.l, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v22, v21 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v7.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v23.l, v19.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v18.h, v13.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v5.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.l, v6.l, v5.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v22 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v22.l, v13.h, v20.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v21.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v21.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v24, v23 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v26.l, v22.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v20.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v6.l, s1 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v4 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v23.l, v19.l, v7.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v26 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v19.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v23.l, v7.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v21.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v12 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v23.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v24, v25 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v4.h, v12.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v22.l, v20.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v20.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v19.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v12.h, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v21 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v6.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v6.l, v20.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v20.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v11 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v22.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v20 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v3.h, v11.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v23.l, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v26, v21 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v19.l, v22.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.l, v11.h, v20.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v20.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v7.l, v5.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v23.l, v21.l -; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v19.l, v5.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v13 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v12 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v20, v20 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v5 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v7.h, v15.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v21, v21 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v16.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.h, v15.h, v16.h, vcc_lo +; GFX12-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v4 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v22, v22 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v26.l, v16.l +; GFX12-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v3 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v16, v17 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v11 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v27.l, v16.l ; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v19.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v7.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v22, v23 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v28.l, v16.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v10.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v25 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.h, v16.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v23, v23 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v24, v24 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v1 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v22.l, v21.l, v20.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v20.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v7.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v10.h, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v23 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v22.l, v20.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v9 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v26.l, v22.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v23 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v23.l, v1.h, v9.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v21.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v26 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.h, v19.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v9.h, v23.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v20.l, v21.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v27, v25 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v23.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v24.l, v5.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.h, v22.l, v19.l, vcc_lo -; GFX12-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v7.l, v6.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v8 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v20.l, v6.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v20.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v21, v24 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v0.h, v8.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v7.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.l, v5.l, v23.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v23.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v8.h, v19.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v16 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v24 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v21.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v21.l, v23.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v23.l, v19.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v24.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v22, v22 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v5.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.l, v16.l, v15.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v22, v22 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v25 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v20.l, v6.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v23, v24 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v15.l, v16.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v22 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v17 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v7.l, v19.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v19.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v24.l, v16.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v22, v22 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v14 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v15.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.l, v14.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v22, v22 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v19.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v25 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v24, v23 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v17.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v21.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v7.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v22 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v17.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v14.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v6.l, v16.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v16.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.l, v16.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v6.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v22, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v18.l, v13.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v24.l, v7.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v6.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.l, v14.l, v17.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v17.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v13.l, v15.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v24 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v16.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v16.l, v17.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v14.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v12 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v17.l, v16.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v17.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l +; GFX12-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff0000, v9 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v6.h, v14.h, s0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v8 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v29, v29 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v18.h, v14.h, v16.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v16.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v7 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v32, v32 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v16, v18 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v14 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v12.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.h, v14.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v18.l, v18.h, v16.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v32, v32 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v13 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.h, v18.l, v16.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v18.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v20, v20 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v32, v32 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v16 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v18, v17 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v11 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v6.l, v15.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v5.h, v13.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v32, v32 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v11 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v25.h, v13.h, v16.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v16.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v32, v32 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v10 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v16, v25 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0, v25.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v32, v32 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v11.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v5.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v15.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v18, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v25.h, v16.h, s1 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v9 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.h, v20.l, v16.h, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v20.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v21, v21 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v32, v32 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v32.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v16 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v3.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v21 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v3.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v5.l, v15.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v17, v17 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v6.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v4.h, v12.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v26.h, v12.h, v16.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v16.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v16, v26 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v10.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v17, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v13 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v12.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v4.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v16.l, v4.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v17, v18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.l, v26.h, v16.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.h, v21.l, v16.h, s3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v21.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v22, v22 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v16 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v13.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v10.l, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v11.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v21, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v8 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v3.h, v11.h, s3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v27.h, v11.h, v16.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v16.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v16, v27 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v9.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v8.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v21, v21 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v12.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v22.l, v27.h, v16.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v22.h, v22.l, v16.h, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v22.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v23, v23 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v16 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v1.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v10.l, v3.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v8.l, v0.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v12.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v22, v17 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v9.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v1.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v0.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v11.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v2.h, v10.h, s4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v28.h, v10.h, v16.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v16.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v16, v28 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.l, v2.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v11.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v18, v17 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v23.l, v28.h, v16.h, s4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v23.h, v23.l, v16.h, s5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v23.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v24, v24 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v16 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v22, v21 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v11 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v9.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v13.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v1.h, v9.h, s5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v29.h, v9.h, v16.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v16.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s5, v16, v29 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v3.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v4.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v8.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v9.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v24.l, v29.h, v16.h, s5 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v24.h, v24.l, v16.h, s6 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v24.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v30, v30 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v30.l, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v0.h, v8.h, s6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.h, v8.h, v16.h, s7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v16.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s6, v16, v30 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v30.h, v16.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v25.l, v15.h, v16.h, s7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v15.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v15.l, s6 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v7.h, s8 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0, v7.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s7, v7, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v16.h, v7.h, s7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v16.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v7.l, v7.h, s8 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v5 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v6.h, v16.h, s7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v7.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v14.l, s8 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v14.l, v6.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0, v6.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v11.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v9.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v14.h, s7 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s8, v6, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v16.h, v6.h, s8 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0, v16.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v6.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v5.h, v16.h, s8 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v6.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.l, v13.l, s9 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v16 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v13.l, v5.h, s10 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0, v5.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v13.h, s8 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s9, v5, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v5.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0, v16.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v5.l, v5.h, s10 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v4.h, v16.h, s9 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v5.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v12.l, s10 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v12.l, v4.h, s11 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0, v4.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v21 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v17 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v12 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.l, v11.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v12.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s10, v4, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v16.h, v4.h, s10 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0, v16.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v4.l, v4.h, s11 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v2 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v3.h, v16.h, s10 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.l, v11.l, s11 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v16 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v11.l, v3.h, s12 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0, v3.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v11.h, s10 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s11, v3, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v16.h, v3.h, s11 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0, v16.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.l, v3.h, s12 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v1 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v2.h, v16.h, s11 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v3.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.l, v10.l, s12 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v10.l, v2.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0, v2.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v2.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v8.l, v1.h, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v10.l, v0.h, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v13.l, v2.h, s3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v0, v16 :: v_dual_mov_b32 v1, v14 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v2, v15 :: v_dual_mov_b32 v3, v20 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v10.h, s11 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s12, v2, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v16.h, v2.h, s12 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0, v16.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v2.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v1.h, v16.h, s12 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v9.l, s13 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v16 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v19 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v9.l, v1.h, s14 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v31, v31 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0, v1.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v8 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v9.h, s12 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s13, v1, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.h, v0.l, v8.l, s14 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0, v16.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v31, v31 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v16.h, v1.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0, v17.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.l, v1.h, s15 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v19.l, v17.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0, v18.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v20.h, v25.h, s17 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0, v32.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v16.h, s16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v1.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v17.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v19.h, v18.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v26.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v20.l, v1.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v8.l, v32.h, s14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v18.l, v0.h, s0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v21.h, v26.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v28.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v27.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v32, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s13 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v21.l, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v29.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v23.h, v28.h, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v16.h, v32.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v16.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v22.h, v27.h, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v24.h, v29.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v30.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v8.l, v32.h, s15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v23.l, v1.h, s4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v22.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v24.l, v8.h, s5 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v25.l, v30.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v9.l, v16.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v8.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v15.h, v0.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v8.h, vcc_lo ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v16bf16: @@ -11681,666 +11539,619 @@ define <32 x bfloat> @v_maximumnum_v32bf16(<32 x bfloat> %x, <32 x bfloat> %y) { ; GFX11-TRUE16-LABEL: v_maximumnum_v32bf16: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff0000, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v30 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v85, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: scratch_load_b32 v68, off, s32 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v15 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff0000, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v30 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v29 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v36.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v36.l ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v33, v33 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v13 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v34, v34 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v13 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff0000, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v14.h, v30.h, s1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v9 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v69, 0xffff0000, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v71, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v86, 0xffff0000, v16 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v55, v55 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s29, v85, v85 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v55.l, v30.h, v32.l, s2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v29 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff0000, v27 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v36.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v12 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v28 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff0000, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v27 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff0000, v26 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v25 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v23 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v68, 0xffff0000, v21 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v70, 0xffff0000, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v80, 0xffff0000, v19 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v81, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v83, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v35, v35 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v37, v37 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v39, v39 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v49, v49 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v51, v51 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v53, v53 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v54, v54 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v65, v65 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v67, v67 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v69, v69 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v71, v71 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s40, v86, v86 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v54.l, v0.h, v16.h, s29 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.l, v32.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.l, v55.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v82, 0xffff0000, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v84, 0xffff0000, v17 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v36, v36 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v38, v38 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v48, v48 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v50, v50 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v52, v52 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s16, v64, v64 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v66, v66 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v68, v68 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v70, v70 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v80, v80 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v81, v81 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v83, v83 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v13.h, v29.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.l, v12.h, v28.h, s5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.l, v11.h, v27.h, s7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.l, v10.h, v26.h, s9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.l, v9.h, v25.h, s11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.l, v8.h, v24.h, s13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.l, v7.h, v23.h, s15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.l, v6.h, v22.h, s17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.l, v5.h, v21.h, s19 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v50.l, v4.h, v20.h, s21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v51.l, v3.h, v19.h, s23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v85.l, v16.h, v54.l, s40 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v86 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v118, 16, v118 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v96, 16, v14 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v82, v82 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v84, v84 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v52.l, v2.h, v18.h, s25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v53.l, v1.h, v17.h, s27 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v64.l, v29.h, v33.l, s4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v65.l, v28.h, v34.l, s6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v66.l, v27.h, v35.l, s8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v67.l, v26.h, v36.l, s10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v68.l, v25.h, v37.l, s12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v69.l, v24.h, v38.l, s14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v70.l, v23.h, v39.l, s16 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v71.l, v22.h, v48.l, s18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v80.l, v21.h, v49.l, s20 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v81.l, v20.h, v50.l, s22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v82.l, v19.h, v51.l, s24 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.l, v54.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s40, v86, v118 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.l, v85.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v97, 16, v30 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v87, v87 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s41, v96, v96 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v87.l, v33.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.l, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v98.l, v36.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v101.l, v39.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v112.l, v50.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v113.l, v51.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v83.l, v18.h, v52.l, s26 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v84.l, v17.h, v53.l, s28 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v64.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.l, v65.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.l, v66.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v67.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v68.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.l, v69.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.l, v70.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v71.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.l, v80.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v81.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, v82.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v116, 16, v116 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v86 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s42, v97, v97 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v30.l, s41 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v32.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v99.l, v37.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v114.l, v52.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v115.l, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v87 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v96, 16, v96 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v98, 16, v98 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v101, 16, v101 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v112, 16, v112 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v113, 16, v113 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v83.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.l, v84.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v118, 16, v119 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v119, 16, v128 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v128, 16, v129 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v129, 16, v130 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v130, 16, v131 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v131, 16, v132 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v132, 16, v133 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v133, 16, v134 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v134, 16, v135 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v135, 16, v144 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v144, 16, v145 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s63, v116, v86 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v86.l, v55.l, v32.l, s40 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v117, 16, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.l, v30.l, v14.l, s42 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0, v55.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v99, 16, v99 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v114, 16, v114 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v115, 16, v115 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v145, 16, v146 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v146, 16, v147 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s42, v87, v118 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s43, v96, v119 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s45, v98, v129 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s56, v101, v132 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s59, v112, v135 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s60, v113, v144 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v86.l, v32.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.l, v86.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.l, v35.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v36.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0, v39.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0, v50.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0, v51.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s46, v99, v130 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s61, v114, v145 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s62, v115, v146 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v96.l, v65.l, v34.l, s43 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v98.l, v67.l, v36.l, s45 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v101.l, v70.l, v39.l, s56 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v112.l, v81.l, v50.l, s59 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v113.l, v82.l, v51.l, s60 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v13.h, v55.l, s16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v118 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v33.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v37.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v102.l, v48.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0, v52.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s14, 0, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v97, 16, v97 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0, v70.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0, v81.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v87.l, v64.l, v33.l, s42 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v99.l, v68.l, v37.l, s46 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v114.l, v83.l, v52.l, s61 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v115.l, v84.l, v53.l, s62 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.h, v96.l, v34.l, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.h, v98.l, v36.l, s5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.l, v101.l, v39.l, s8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v101.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.h, v112.l, v50.l, s11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v112.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.l, v113.l, v51.l, s12 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v113.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v100.l, v38.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v103.l, v49.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v102, 16, v102 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s44, v97, v128 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v116.l, v85.l, v54.l, s63 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v87.l, v33.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v99.l, v37.l, s6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v114.l, v52.l, s13 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v114.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.l, v115.l, v53.l, s14 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v115.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.l, v34.l, v70.l, s23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v70, 16, v39 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.l, v35.h, v81.l, s26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v35.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v100, 16, v100 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v103, 16, v103 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s57, v102, v133 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v97.l, v66.l, v35.l, s44 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.h, v116.l, v54.l, s15 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v116.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v51 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0, v48.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0, v65.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0, v66.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s47, v100, v131 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s58, v103, v134 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v102.l, v71.l, v48.l, s57 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.l, v96.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v97.l, v35.l, s4 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.l, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v52 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s14, 0, v53 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v38.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0, v49.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0, v64.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0, v71.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v100.l, v69.l, v38.l, s47 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v103.l, v80.l, v49.l, s58 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v87.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.h, v102.l, v48.l, s9 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v102.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.h, v30.h, v65.l, s18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v128 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v32.l, v66.l, s19 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 16, v129 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0, v80.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s41, 0, v85.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.h, v100.l, v38.l, s7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.l, v103.l, v49.l, s10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v103.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.l, v14.h, v64.l, s17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 16, v119 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.h, v34.h, v71.l, s24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v71, 16, v48 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v65 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v66 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v86.l, v13.h, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.l, v35.l, v80.l, s25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 16, v49 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.l, v37.h, v85.l, s41 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v64 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v71 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.h, v96.l, v30.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.h, v97.l, v32.l, s4 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0, v67.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0, v68.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v98.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v99.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v87.l, v38.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.h, v102.l, v38.h, s9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.h, v32.h, v67.l, s20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v67, 16, v130 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v33.l, v68.l, s21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v68, 16, v131 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s29, 0, v84.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v70 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v67 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s27, 0, v82.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v68 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.l, v37.l, v84.l, s29 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.h, v101.l, v34.l, s8 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0, v83.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v80 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.l, v36.l, v82.l, s27 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.h, v115.l, v37.l, s14 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0, v69.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.l, v36.h, v83.l, s28 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v103.l, v35.l, s10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.h, v113.l, v36.l, s12 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.l, v100.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.h, v33.h, v69.l, s22 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v81 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.h, v114.l, v48.l, s13 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 16, v132 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v69 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.h, v100.l, v33.h, s7 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v25 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v39, v39 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v49, v49 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v10 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v50, v50 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff0000, v9 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v51, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v8 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v23 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v22 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v21 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v20 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v52, v52 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v7 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v53, v53 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v54, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v55, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v64, v64 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v19 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v18 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v17 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v81, 0xffff0000, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v65, v65 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v66, v66 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v67, v67 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v15 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v83, 16, v30 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v84, 16, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v85, 16, v17 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v16 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v83, v83 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s42, v86, v86 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v96.h, v0.l, v16.l, s42 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s43, 0, v96.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v31 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v50.l, v15.h, v31.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v31.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v55, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v68 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v15.h, v68.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v36.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.h, v68.h, v36.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v36, v35 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s42, 0, v35.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.l, v35.h, v36.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.h, v31.l, v36.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v31.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v32, v32 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v14.h, v30.h, s0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.h, v30.h, v36.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v36, v37 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s44, 0, v37.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v37.h, v36.h, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.h, v32.l, v36.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v32.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v33, v33 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v13.h, v29.h, s1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.h, v29.h, v36.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v36, v38 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v38.h, v36.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.h, v33.l, v36.h, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v33.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v34, v34 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v12.h, v28.h, s2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.h, v28.h, v36.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v36, v39 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.l, v39.h, v36.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.h, v34.l, v36.h, s3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v34.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v48, v48 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v36.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v50.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v15.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v51.l, v31.h, v50.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v31 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.h, v98.l, v32.h, s5 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v51.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.h, v99.l, v33.l, s6 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v54, v54 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.l, v31.l, v15.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v50.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v52, v53 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v31.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v11.h, v27.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.h, v27.h, v36.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v36, v48 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.l, v48.h, v36.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.l, v35.l, v36.h, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v35.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v49, v49 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v36.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v51.l, v50.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v52 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v51.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v32.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v32.l, v50.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v10.h, v26.h, s4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.h, v26.h, v36.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v36, v49 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.l, v49.h, v36.h, s4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.l, v38.l, v36.h, s5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v38.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v50, v50 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v9.h, v25.h, s5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v50.h, v25.h, v36.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s5, v36, v50 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.l, v50.h, v36.h, s5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.l, v48.l, v36.h, s6 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v48.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v51, v51 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v8.h, v24.h, s6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v51.h, v24.h, v36.h, s7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s6, v36, v51 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v50.l, v51.h, v36.h, s6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v51.l, v50.l, v36.h, s7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v50.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v52, v52 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v7.h, v23.h, s7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v52.h, v23.h, v36.h, s8 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s7, v36, v52 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v52.l, v52.h, v36.h, s7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v69.l, v52.l, v36.h, s8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v52.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v53, v53 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v6.h, v22.h, s8 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v53.h, v22.h, v36.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s8, v36, v53 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v53.l, v53.h, v36.h, s8 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v69.h, v53.l, v36.h, s9 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v53.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v54, v54 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v5.h, v21.h, s9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v54.h, v21.h, v36.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s9, v36, v54 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v54.l, v54.h, v36.h, s9 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v70.l, v54.l, v36.h, s10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v54.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v55, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v4.h, v20.h, s10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v55.h, v20.h, v36.h, s11 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s10, v36, v55 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v55.l, v55.h, v36.h, s10 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v70.h, v55.l, v36.h, s11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v55.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v64, v64 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v3.h, v19.h, s11 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v64.h, v19.h, v36.h, s12 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s11, v36, v64 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v64.l, v64.h, v36.h, s11 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v71.l, v64.l, v36.h, s12 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v64.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v65, v65 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v2.h, v18.h, s12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v65.h, v18.h, v36.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s12, v36, v65 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v65.l, v65.h, v36.h, s12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v71.h, v65.l, v36.h, s13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v65.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v66, v66 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v1.h, v17.h, s13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v66.h, v17.h, v36.h, s14 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s14, 0, v36.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s13, v36, v66 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v66.l, v66.h, v36.h, s13 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v80.l, v66.l, v36.h, s14 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v66.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v67, v67 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v0.h, v16.h, s14 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v67.h, v16.h, v36.h, s15 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0, v36.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v68 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s14, v36, v67 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v82.h, v15.l, v68.l, s15 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v14 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v67.l, v67.h, v36.h, s14 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v67.l, v36.h, s16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v67.l +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0, v82.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s14, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v68.l, v82.h, s15 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s15, v82, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v68.l, v36.h, v82.h, s15 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0, v36.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v68.l, v82.h, s16 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s16, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v13 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v29 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.h, v14.h, v36.h, s15 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v68.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v14.l, v30.l, s16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v28 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s15, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v30.l, v14.h, s17 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0, v14.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v38, v53 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.h, v112.l, v39.l, s11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v52 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v15.h, v51.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v31.l, v15.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v29 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v50 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v30.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v33.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v33.l, v15.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v32.l, v15.h, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v117, v117 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v31.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v29.h, v116.l, v49.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v27 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s16, v14, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v82, v82 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v36.h, v14.h, s16 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0, v36.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v26 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v14.l, v14.h, s17 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v12 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v25 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v29.h, v13.h, v36.h, s16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v14.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v13.l, v29.l, s17 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v24 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s16, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v29.l, v13.h, s18 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0, v13.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v23 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s17, v13, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v29.h, s16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v82, v82 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v36.h, v13.h, s17 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0, v36.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v22 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v13.l, v13.h, s18 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v11 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v21 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v28.h, v12.h, v36.h, s17 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v13.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v12.l, v28.l, s18 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v20 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s17, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v28.l, v12.h, s19 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0, v12.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v19 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s18, v12, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v13.l, v28.h, s17 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v82, v82 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v36.h, v12.h, s18 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0, v36.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v12.l, v12.h, s19 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v10 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v27.h, v11.h, v36.h, s18 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v12.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v11.l, v27.l, s19 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s18, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v27.l, v11.h, s20 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0, v11.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v27.h, s18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s19, v11, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v36.h, v11.h, s19 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v11.l, v11.h, s20 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v9 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v26.h, v10.h, v36.h, s19 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v11.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.l, v26.l, s20 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s19, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v26.l, v10.h, s21 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0, v10.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v26.h, s19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s20, v10, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v36.h, v10.h, s20 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v10.l, v10.h, s21 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v25.h, v9.h, v36.h, s20 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v10.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.l, v25.l, s21 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s20, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v25.l, v9.h, s22 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0, v9.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v10.l, v25.h, s20 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s21, v9, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v36.h, v9.h, s21 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v9.l, v9.h, s22 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v24.h, v8.h, v36.h, s21 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.l, v24.l, s22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s21, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v24.l, v8.h, s23 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0, v8.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v24.h, s21 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s22, v8, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v36.h, v8.h, s22 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v8.l, v8.h, s23 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v23.h, v7.h, v36.h, s22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v8.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v23.l, s23 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s22, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v23.l, v7.h, s24 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0, v7.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v23.h, s22 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s23, v7, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v36.h, v7.h, s23 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v7.l, v7.h, s24 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v22.h, v6.h, v36.h, s23 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v22.l, s24 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s23, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v22.l, v6.h, s25 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v22.h, s23 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s24, v6, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v36.h, v6.h, s24 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v6.h, s25 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.h, v5.h, v36.h, s24 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.l, v21.l, s25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s24, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v21.l, v5.h, s26 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0, v5.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v21.h, s24 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s25, v5, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v36.h, v5.h, s25 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v5.l, v5.h, s26 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.h, v4.h, v36.h, s25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v5.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v20.l, s26 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s25, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v20.l, v4.h, s27 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s27, 0, v4.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v20.h, s25 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s26, v4, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v36.h, v4.h, s26 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v4.l, v4.h, s27 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v18 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v18.h, v3.h, v36.h, s26 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v4.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v83.h, v3.l, v19.l, s27 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s29, v81, v81 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s26, 0, v36 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v13.l, v29.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v52, v52 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v53 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v51, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v31.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v29.l, v29.l, v13.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v52 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.l, v30.l, v14.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v13.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v28 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v31.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v28.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v31.l, v14.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v53 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v52, v51 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v28.l, v28.l, v12.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v12.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v12.h, v30.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v50 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.l, v29.l, v13.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v33.l, v15.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v13.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v31.l, v12.h, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v27 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v30.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v30.l, v13.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v51, v50 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v27.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v53 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v29.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.l, v28.l, v12.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v27.l, v27.l, v11.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v11.h, v29.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v31.l, v12.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v11.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v26 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v28.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v31.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v10.l, v26.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v52 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v30.l, v11.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v51, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.h, v28.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v26.l, v10.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v50 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v26.l, v27.l, v11.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v28, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v25.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v28, v28 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v26.l, v11.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v52 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v51, v50 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v25.l, v9.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.l, v31.l, v10.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.h, v27.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v28 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v25.l, v12.l, v10.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v9.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.l, v26.l, v9.h, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v27 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v28 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v25.l, v10.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v50, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v24 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v27, v26 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v12.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v24.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v28 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v11.l, v9.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v9.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.h, v12.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v24.l, v8.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v7 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v26 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v10.l, v9.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v12.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v24, v24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v23 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v23.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v24, v24 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v27 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.l, v25.l, v8.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v26, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v23.l, v7.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.h, v11.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v12.l, v8.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v22 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v22.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v11.l, v8.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v26 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v25, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v22.l, v6.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v10.l, v7.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.h, v12.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v9.l, v7.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v6.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v11.l, v6.h, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v10.l, v7.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v21 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v10.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v12, v11 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v9.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v21.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v8.l, v6.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v9.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v21.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v7.l, v6.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v20.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.l, v10.l, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v20.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.h, v8.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v9.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v19 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v19.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v8.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v20 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v12, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v19.l, v3.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v9.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v6.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v18.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v9.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.l, v7.l, v4.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v18.l, v2.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v11, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.l, v8.l, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v6.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v17 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v16 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v17.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v16.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v17.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v3.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v16.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v12, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v3.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.l, v2.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v17, v16 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v7.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v4.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v8.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v7.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v16 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v11 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v29.l, v5.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.l, v2.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.l, v8.l, v1.h, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.l, v6.l, v0.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.l, v9.l, v2.h, s3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, v29 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v49 :: v_dual_mov_b32 v2, v48 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v39 :: v_dual_mov_b32 v4, v38 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v36 :: v_dual_mov_b32 v6, v35 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, v34 :: v_dual_mov_b32 v8, v33 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v32 :: v_dual_mov_b32 v10, v31 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, v30 :: v_dual_mov_b32 v12, v37 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v19.l, v83.h, s28 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v84, v84 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s40, v3, v3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s41, 0, v83.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s27, v83, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v82.h, v2.l, v18.l, s28 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0, v36.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.h, v1.l, v17.l, s40 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v32.h, v37.h, s44 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v36.h, v83.h, s27 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s40, 0, v82.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v85, v85 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v18.h, s26 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v32.l, v1.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v83.h, s41 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s41, v87, v87 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.h, v36.h, s28 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v3.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v31.h, v35.h, s42 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s42, 0, v39.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0, v19.h +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s45, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v18.l, v82.h, s29 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s29, 0, v38.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v31.l, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v34.h, v39.h, s42 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v48.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s44, v82, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v33.h, v38.h, s29 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v36.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v34.l, v1.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v37.l, v48.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v36.h, v82.h, s44 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v33.l, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v49.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v50.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v35.l, v1.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v82.h, s40 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v51.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v54.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v49.l, v50.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.l, s45 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v36.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v39.l, v49.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v52.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v51.l, v51.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v53.h +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v17.l, v19.h, s27 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v38.l, v0.h, s4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v69.l, v52.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v55.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v48.l, v1.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v19, v36 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v36.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v69.h, v53.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v70.h, v55.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v66.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v36.h, v19.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v65.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v50.l, v2.h, s6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v70.l, v54.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v52.l, v0.h, s7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.l, v16.h, v19.h, s28 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v53.l, v1.h, s8 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v71.h, v65.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v54.l, v2.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v67.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.l, v36.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v16.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v64.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v65.l, v1.h, s12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v67.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v55.l, v3.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v16.l, v96.h, s41 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v71.l, v64.h, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.l, v80.l, v66.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v2.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v16.h, v17.l, s4 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v96, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v64.l, v0.h, s11 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v66.l, v16.l, s13 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v67.l, v15.l, s14 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v68.l, v30.h, s15 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.l, v36.h, v96.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.h, v16.l, v96.h, s43 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.h, v17.h, v36.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v16.l, v17.h, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v32bf16: @@ -12956,753 +12767,697 @@ define <32 x bfloat> @v_maximumnum_v32bf16(<32 x bfloat> %x, <32 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: scratch_load_b32 v31, off, s32 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v14 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff0000, v15 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v30 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v7 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v85, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: scratch_load_b32 v68, off, s32 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v15 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff0000, v14 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v30 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v29 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v35.l, v36.l +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v37.l, v36.l ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v33, v33 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v13 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v34, v34 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v13 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff0000, v12 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v14.h, v30.h, s1 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v11 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v10 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v9 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v8 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v24 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v6 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v5 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v69, 0xffff0000, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v71, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v86, 0xffff0000, v16 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v55, v55 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s29, v85, v85 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v55.l, v30.h, v32.l, s2 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v29 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v28 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff0000, v27 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v38.l, v36.l +; GFX12-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v12 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v28 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff0000, v11 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v27 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff0000, v26 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v25 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v23 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v22 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v68, 0xffff0000, v21 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v70, 0xffff0000, v20 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v80, 0xffff0000, v19 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v81, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v83, 0xffff0000, v1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v35, v35 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v37, v37 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v39, v39 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v49, v49 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v51, v51 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v53, v53 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v54, v54 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v65, v65 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v67, v67 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v69, v69 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v71, v71 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s40, v86, v86 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v54.l, v0.h, v16.h, s29 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v86.l, v32.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v118.l, v55.l -; GFX12-TRUE16-NEXT: v_and_b32_e32 v82, 0xffff0000, v18 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v84, 0xffff0000, v17 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v36, v36 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v38, v38 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v48, v48 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v50, v50 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v52, v52 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s16, v64, v64 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v66, v66 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v68, v68 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v70, v70 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v80, v80 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v81, v81 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v83, v83 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v13.h, v29.h, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.l, v12.h, v28.h, s5 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.l, v11.h, v27.h, s7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.l, v10.h, v26.h, s9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.l, v9.h, v25.h, s11 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.l, v8.h, v24.h, s13 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.l, v7.h, v23.h, s15 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.l, v6.h, v22.h, s17 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.l, v5.h, v21.h, s19 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v50.l, v4.h, v20.h, s21 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v51.l, v3.h, v19.h, s23 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v85.l, v16.h, v54.l, s40 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v86 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v118, 16, v118 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v15 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v96, 16, v14 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v82, v82 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v84, v84 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v52.l, v2.h, v18.h, s25 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v53.l, v1.h, v17.h, s27 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v64.l, v29.h, v33.l, s4 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v65.l, v28.h, v34.l, s6 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v66.l, v27.h, v35.l, s8 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v67.l, v26.h, v36.l, s10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v68.l, v25.h, v37.l, s12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v69.l, v24.h, v38.l, s14 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v70.l, v23.h, v39.l, s16 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v71.l, v22.h, v48.l, s18 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v80.l, v21.h, v49.l, s20 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v81.l, v20.h, v50.l, s22 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v82.l, v19.h, v51.l, s24 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v116.l, v54.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s40, v86, v118 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v86.l, v85.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v97, 16, v30 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v87, v87 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s41, v96, v96 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v87.l, v33.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v96.l, v34.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v98.l, v36.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v101.l, v39.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v112.l, v50.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v113.l, v51.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v83.l, v18.h, v52.l, s26 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v84.l, v17.h, v53.l, s28 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v119.l, v64.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v128.l, v65.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v129.l, v66.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v130.l, v67.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v131.l, v68.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v132.l, v69.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v133.l, v70.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v134.l, v71.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v135.l, v80.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v144.l, v81.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v145.l, v82.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v116, 16, v116 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v86 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s42, v97, v97 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v30.l, s41 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v32.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v99.l, v37.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v114.l, v52.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v115.l, v53.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v87 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v96, 16, v96 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v98, 16, v98 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v101, 16, v101 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v112, 16, v112 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v113, 16, v113 -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v146.l, v83.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v147.l, v84.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v118, 16, v119 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v119, 16, v128 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v128, 16, v129 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v129, 16, v130 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v130, 16, v131 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v131, 16, v132 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v132, 16, v133 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v133, 16, v134 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v134, 16, v135 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v135, 16, v144 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v144, 16, v145 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s63, v116, v86 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v86.l, v55.l, v32.l, s40 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v117, 16, v13 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.l, v30.l, v14.l, s42 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0, v55.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v99, 16, v99 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v114, 16, v114 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v115, 16, v115 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v145, 16, v146 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v146, 16, v147 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s42, v87, v118 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s43, v96, v119 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s45, v98, v129 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s56, v101, v132 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s59, v112, v135 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s60, v113, v144 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v86.l, v32.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v118.l, v86.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v34.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v97.l, v35.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v36.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0, v39.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0, v50.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0, v51.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s46, v99, v130 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s61, v114, v145 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s62, v115, v146 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v96.l, v65.l, v34.l, s43 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v98.l, v67.l, v36.l, s45 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v101.l, v70.l, v39.l, s56 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v112.l, v81.l, v50.l, s59 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v113.l, v82.l, v51.l, s60 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v13.h, v55.l, s16 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v118 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v33.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v37.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v102.l, v48.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0, v52.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s14, 0, v53.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v97, 16, v97 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0, v70.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0, v81.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v87.l, v64.l, v33.l, s42 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v99.l, v68.l, v37.l, s46 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v114.l, v83.l, v52.l, s61 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v115.l, v84.l, v53.l, s62 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.h, v96.l, v34.l, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.h, v98.l, v36.l, s5 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.l, v101.l, v39.l, s8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v39.l, v101.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.h, v112.l, v50.l, s11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v50.l, v112.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.l, v113.l, v51.l, s12 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v113.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v55 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v100.l, v38.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v103.l, v49.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0, v54.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v102, 16, v102 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s44, v97, v128 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v116.l, v85.l, v54.l, s63 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v87.l, v33.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v99.l, v37.l, s6 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v114.l, v52.l, s13 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v114.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.l, v115.l, v53.l, s14 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v115.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.l, v34.l, v70.l, s23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v70, 16, v39 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.l, v35.h, v81.l, s26 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v35.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v100, 16, v100 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v103, 16, v103 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s57, v102, v133 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v97.l, v66.l, v35.l, s44 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.h, v116.l, v54.l, s15 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v54.l, v116.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v51 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0, v48.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0, v65.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0, v66.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s47, v100, v131 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s58, v103, v134 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v102.l, v71.l, v48.l, s57 -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v128.l, v96.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v97.l, v35.l, s4 -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v129.l, v97.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v52 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s14, 0, v53 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v38.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0, v49.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0, v64.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0, v71.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v100.l, v69.l, v38.l, s47 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v103.l, v80.l, v49.l, s58 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v119.l, v87.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.h, v102.l, v48.l, s9 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v48.l, v102.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.h, v30.h, v65.l, s18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v128 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v32.l, v66.l, s19 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 16, v129 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0, v80.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s41, 0, v85.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.h, v100.l, v38.l, s7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.l, v103.l, v49.l, s10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v49.l, v103.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.l, v14.h, v64.l, s17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 16, v119 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.h, v34.h, v71.l, s24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v71, 16, v48 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v65 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v66 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v86.l, v13.h, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.l, v35.l, v80.l, s25 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 16, v49 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.l, v37.h, v85.l, s41 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v64 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v71 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.h, v96.l, v30.h, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.h, v97.l, v32.l, s4 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0, v67.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0, v68.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v130.l, v98.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v131.l, v99.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v87.l, v38.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.h, v102.l, v38.h, s9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.h, v32.h, v67.l, s20 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v67, 16, v130 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v33.l, v68.l, s21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v68, 16, v131 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s29, 0, v84.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v70 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v67 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s27, 0, v82.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v68 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.l, v37.l, v84.l, s29 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.h, v101.l, v34.l, s8 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0, v83.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v80 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.l, v36.l, v82.l, s27 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.h, v115.l, v37.l, s14 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0, v69.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.l, v36.h, v83.l, s28 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v103.l, v35.l, s10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.h, v113.l, v36.l, s12 -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v132.l, v100.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.h, v33.h, v69.l, s22 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v81 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.h, v114.l, v48.l, s13 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 16, v132 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v69 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.h, v100.l, v33.h, s7 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v25 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v39, v39 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v39.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v49, v49 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v10 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v50, v50 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff0000, v9 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v51, v51 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v8 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v24 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v23 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v22 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v21 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v20 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v52, v52 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v7 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v53, v53 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v6 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v54, v54 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v55, v55 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v4 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v64, v64 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v19 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v18 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v17 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v81, 0xffff0000, v16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v82.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v65, v65 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v66, v66 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v67, v67 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v15 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v83, 16, v30 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v84, 16, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v0 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v85, 16, v17 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v16 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v83, v83 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v83.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s42, v86, v86 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v96.l, v36.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v96.h, v0.l, v16.l, s42 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s43, 0, v96.h ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v31 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v68 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v50.l, v15.h, v31.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v31.l, s0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v55, v55 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v50.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v15.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v15.h, v68.h, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v36.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v51.l, v31.h, v50.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v54 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v31 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.h, v98.l, v32.h, s5 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v51.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.h, v99.l, v33.l, s6 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v54, v54 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.l, v31.l, v15.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v50.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v52, v53 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v31.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v51.l, v50.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v52 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v51.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v32.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v32.l, v50.l, s0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v38, v53 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.h, v112.l, v39.l, s11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v52 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v15.h, v51.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v31.l, v15.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v14.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v29 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v50 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v50.l, v30.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v33.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v33.l, v15.l, s2 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v32.l, v15.h, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v117, v117 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v31.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.h, v68.h, v36.h, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v36, v35 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s42, 0, v35.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v29.h, v116.l, v49.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.l, v35.h, v36.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v13.l, v29.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v52, v52 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v53 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v51, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v31.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v29.l, v29.l, v13.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v52 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.l, v30.l, v14.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v13.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v29.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v28 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v14.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v31.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.h, v31.l, v36.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v31.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v32, v32 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v14.h, v30.h, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.h, v30.h, v36.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v36, v37 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s44, 0, v37.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v37.h, v36.h, s0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.h, v32.l, v36.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v32.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v33, v33 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v28.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v31.l, v14.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v30.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v53 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v52, v51 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v13.h, v29.h, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.h, v29.h, v36.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v36, v38 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v28.l, v28.l, v12.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v12.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v12.h, v30.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v50 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.l, v29.l, v13.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v50.l, v28.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v11 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v33.l, v15.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v13.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v31.l, v12.h, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v27 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v30.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v38.h, v36.h, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.h, v33.l, v36.h, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v33.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v34, v34 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v30.l, v13.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v51, v50 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v27.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v53 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v29.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v12.h, v28.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.h, v28.h, v36.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v36, v39 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.l, v28.l, v12.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v12.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v27.l, v27.l, v11.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v11.h, v29.l, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.l, v39.h, v36.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.h, v34.l, v36.h, s3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v34.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v48, v48 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v48.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v31.l, v12.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v27.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v11.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v26 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v28.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v31.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v11.h, v27.h, s3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.h, v27.h, v36.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v36, v48 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v10.l, v26.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v52 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v30.l, v11.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v51, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.h, v28.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.l, v48.h, v36.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.l, v35.l, v36.h, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v35.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v49, v49 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v49.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v26.l, v10.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v50 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v26.l, v27.l, v11.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v10.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v28, v28 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v11.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v26.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v10.h, v26.h, s4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.h, v26.h, v36.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s4, v36, v49 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v25.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v28, v28 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v26.l, v11.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v27.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v52 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v51, v50 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.l, v49.h, v36.h, s4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.l, v38.l, v36.h, s5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v38.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v50, v50 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v50.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v25.l, v9.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.l, v31.l, v10.h, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.h, v27.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v28 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v25.l, v12.l, v10.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v27.l, v11.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v28.l, v9.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v10.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v8 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.l, v26.l, v9.h, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v27 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v28 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v9.h, v25.h, s5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v50.h, v25.h, v36.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s5, v36, v50 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v25.l, v10.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v50, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v24 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v27, v26 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v12.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v24.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v28 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.l, v50.h, v36.h, s5 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.l, v48.l, v36.h, s6 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v48.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v51, v51 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v11.l, v9.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v9.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.h, v12.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v24.l, v8.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v7 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v26 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v8.h, v24.h, s6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v51.h, v24.h, v36.h, s7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s6, v36, v51 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v10.l, v9.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v26.l, v8.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v12.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v24, v24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v23 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v27.l, v10.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v26 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v50.l, v51.h, v36.h, s6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v51.l, v50.l, v36.h, s7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v50.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v52, v52 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v23.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v24, v24 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v11.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v27 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.l, v25.l, v8.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v26, v9 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v7.h, v23.h, s7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v52.h, v23.h, v36.h, s8 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s7, v36, v52 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v23.l, v7.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v6 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.h, v11.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v24 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v12.l, v8.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v24.l, v9.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v22 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v8.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v26.l, v11.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v52.l, v52.h, v36.h, s7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v69.l, v52.l, v36.h, s8 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v52.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v53, v53 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v22.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v11.l, v8.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v12.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v26 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v25, v24 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v6.h, v22.h, s8 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v53.h, v22.h, v36.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s8, v36, v53 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v22.l, v6.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v10.l, v7.h, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.h, v12.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v23 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v9.l, v7.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v5 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v11.l, v6.h, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v22 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v10.l, v7.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v21 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v10.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v12, v11 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v9.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v21.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v22 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v53.l, v53.h, v36.h, s8 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v69.h, v53.l, v36.h, s9 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v53.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v54, v54 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v54.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v8.l, v6.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v9.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v21.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v5.h, v21.h, s9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v54.h, v21.h, v36.h, s10 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s9, v36, v54 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v7.l, v6.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v20 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v8.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v54.l, v54.h, v36.h, s9 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v70.l, v54.l, v36.h, s10 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v54.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v55, v55 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v55.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v20.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.l, v10.l, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v4.h, v20.h, s10 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v55.h, v20.h, v36.h, s11 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s10, v36, v55 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v20.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.h, v8.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v11 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v9.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v4.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v19 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v20.l, v8.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v55.l, v55.h, v36.h, s10 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v70.h, v55.l, v36.h, s11 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v55.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v64, v64 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v64.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v19.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v8.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v9.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v20 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v12, v11 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v3.h, v19.h, s11 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v64.h, v19.h, v36.h, s12 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s11, v36, v64 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v19.l, v3.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v9.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v18 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v6.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v4.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v64.l, v64.h, v36.h, s11 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v71.l, v64.l, v36.h, s12 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v64.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v65, v65 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v65.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v18.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v9.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v6.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.l, v7.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v2.h, v18.h, s12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v65.h, v18.h, v36.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s12, v36, v65 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v18.l, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v11, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.l, v8.l, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v6.l, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v65.l, v65.h, v36.h, s12 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v71.h, v65.l, v36.h, s13 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v65.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v66, v66 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v66.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v17 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v16 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v9.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v1.h, v17.h, s13 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v66.h, v17.h, v36.h, s14 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s14, 0, v36.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s13, v36, v66 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v17.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v16.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v66.l, v66.h, v36.h, s13 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v80.l, v66.l, v36.h, s14 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v66.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v67, v67 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v67.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v0.h, v16.h, s14 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v67.h, v16.h, v36.h, s15 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0, v36.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v68 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s14, v36, v67 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v82.h, v15.l, v68.l, s15 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v67.l, v67.h, v36.h, s14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v67.l, v36.h, s16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v67.l +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0, v82.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s14, 0, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v68.l, v82.h, s15 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s15, v82, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v68.l, v36.h, v82.h, s15 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v68.l, v82.h, s16 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s16, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v13 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v29 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.h, v14.h, v36.h, s15 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v68.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v14.l, v30.l, s16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v28 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s15, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v30.l, v14.h, s17 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0, v14.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v27 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s16, v14, v36 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v82, v82 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v36.h, v14.h, s16 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0, v36.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v26 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v14.l, v14.h, s17 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v12 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v25 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v29.h, v13.h, v36.h, s16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v14.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v13.l, v29.l, s17 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v24 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s16, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v29.l, v13.h, s18 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0, v13.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v23 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s17, v13, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v17.l, v1.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v3.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v16.l, v0.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v12, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v1.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v7.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, v3.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v5.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v29.h, s16 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v82, v82 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v36.h, v13.h, s17 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0, v36.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v22 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v13.l, v13.h, s18 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v11 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v21 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v28.h, v12.h, v36.h, s17 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v13.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v12.l, v28.l, s18 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v20 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s17, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v28.l, v12.h, s19 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0, v12.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v19 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s18, v12, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.l, v2.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v12, v11 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v13.l, v28.h, s17 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v82, v82 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v36.h, v12.h, s18 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0, v36.h +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v82.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v12.l, v12.h, s19 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v10 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v27.h, v11.h, v36.h, s18 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v12.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v11.l, v27.l, s19 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s18, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v27.l, v11.h, s20 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0, v11.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v17, v16 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v5 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v7.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v27.h, s18 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s19, v11, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v36.h, v11.h, s19 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v11.l, v11.h, s20 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v9 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v26.h, v10.h, v36.h, s19 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v11.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.l, v26.l, s20 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s19, 0, v36 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v26.l, v10.h, s21 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0, v10.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v26.h, s19 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s20, v10, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v36.h, v10.h, s20 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0, v36.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v10.l, v10.h, s21 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v8 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v25.h, v9.h, v36.h, s20 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v10.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.l, v25.l, s21 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s20, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v25.l, v9.h, s22 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0, v9.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v4.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v8.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, v2.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v5.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v10.l, v25.h, s20 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s21, v9, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v36.h, v9.h, s21 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v9.l, v9.h, s22 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v7 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v24.h, v8.h, v36.h, s21 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v9.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.l, v24.l, s22 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s21, 0, v36 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v24.l, v8.h, s23 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0, v8.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v24.h, s21 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s22, v8, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v36.h, v8.h, s22 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0, v36.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v8.l, v8.h, s23 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v23.h, v7.h, v36.h, s22 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v8.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v23.l, s23 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s22, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v23.l, v7.h, s24 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0, v7.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v7.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v23.h, s22 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s23, v7, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v36.h, v7.h, s23 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v7.l, v7.h, s24 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v5 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v22.h, v6.h, v36.h, s23 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v7.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v22.l, s24 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s23, 0, v36 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v22.l, v6.h, s25 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0, v6.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v22.h, s23 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s24, v6, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v36.h, v6.h, s24 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0, v36.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v6.h, s25 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v16 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v11 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v10 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v29.l, v5.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.h, v5.h, v36.h, s24 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v6.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.l, v21.l, s25 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s24, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v21.l, v5.h, s26 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0, v5.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.l, v2.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.l, v8.l, v1.h, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.l, v6.l, v0.h, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.l, v9.l, v2.h, s3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v0, v29 :: v_dual_mov_b32 v1, v49 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v2, v48 :: v_dual_mov_b32 v3, v39 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v4, v38 :: v_dual_mov_b32 v5, v36 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v6, v35 :: v_dual_mov_b32 v7, v34 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v8, v33 :: v_dual_mov_b32 v9, v32 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v10, v31 :: v_dual_mov_b32 v11, v30 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v12, v37 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v21.h, s24 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s25, v5, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v36.h, v5.h, s25 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v5.l, v5.h, s26 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.h, v4.h, v36.h, s25 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v5.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v20.l, s26 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s25, 0, v36 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v20.l, v4.h, s27 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s27, 0, v4.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v20.h, s25 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s26, v4, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v36.h, v4.h, s26 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0, v36.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v4.l, v4.h, s27 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v18 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v18.h, v3.h, v36.h, s26 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v4.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v83.h, v3.l, v19.l, s27 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s29, v81, v81 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s26, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v19.l, v83.h, s28 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v84, v84 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s40, v3, v3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s41, 0, v83.h +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v19.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s27, v83, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v82.h, v2.l, v18.l, s28 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.h, v1.l, v17.l, s40 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v32.h, v37.h, s44 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v36.h, v83.h, s27 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s40, 0, v82.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v85, v85 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v18.h, s26 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v32.l, v1.l, s0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v83.h, s41 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s41, v87, v87 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.h, v36.h, s28 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v3.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v31.h, v35.h, s42 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s42, 0, v39.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0, v19.h +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s45, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v18.l, v82.h, s29 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s29, 0, v38.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v31.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v34.h, v39.h, s42 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v48.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s44, v82, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v33.h, v38.h, s29 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v34.l, v1.h, s2 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v37.l, v48.h, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v36.h, v82.h, s44 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v33.l, v0.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v49.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v50.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v35.l, v1.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v82.h, s40 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v51.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v54.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v49.l, v50.h, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.l, s45 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v36.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v1.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v39.l, v49.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v52.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v51.l, v51.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v53.h +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v17.l, v19.h, s27 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v38.l, v0.h, s4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v69.l, v52.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v55.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v48.l, v1.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v19, v36 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v69.h, v53.h, s2 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v70.h, v55.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v66.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v36.h, v19.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v65.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v50.l, v2.h, s6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v70.l, v54.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v52.l, v0.h, s7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.l, v16.h, v19.h, s28 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v53.l, v1.h, s8 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v71.h, v65.h, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v54.l, v2.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v67.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.l, v36.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v16.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v64.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v65.l, v1.h, s12 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v67.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v55.l, v3.h, s10 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v16.l, v96.h, s41 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v71.l, v64.h, s0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.l, v80.l, v66.h, s2 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v2.l, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v16.h, v17.l, s4 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v96, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v64.l, v0.h, s11 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v66.l, v16.l, s13 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v67.l, v15.l, s14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v68.l, v30.h, s15 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.l, v36.h, v96.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v36.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.h, v16.l, v96.h, s43 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.h, v17.h, v36.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v16.l, v17.h, s0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v32bf16: @@ -14612,34 +14367,29 @@ define bfloat @v_maximumnum_bf16_no_ieee(bfloat %x, bfloat %y) #0 { ; GFX11-TRUE16-LABEL: v_maximumnum_bf16_no_ieee: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v1, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_bf16_no_ieee: @@ -14675,40 +14425,34 @@ define bfloat @v_maximumnum_bf16_no_ieee(bfloat %x, bfloat %y) #0 { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v2, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v1.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v1, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v1.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_bf16_no_ieee: @@ -14949,58 +14693,46 @@ define <2 x bfloat> @v_maximumnum_v2bf16_no_ieee(<2 x bfloat> %x, <2 x bfloat> % ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v5, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v4.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v4, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.h, v4.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v4.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v1.l, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.l, v3.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v3, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v3.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v3.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v2.l, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v4, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v1.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v2bf16_no_ieee: @@ -15057,62 +14789,56 @@ define <2 x bfloat> @v_maximumnum_v2bf16_no_ieee(<2 x bfloat> %x, <2 x bfloat> % ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v5, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v1.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v4, v6 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v4.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v4, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.h, v4.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v1.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v1.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.l, v3.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v3, v4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v3.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v3.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v2.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v2bf16_no_ieee: @@ -15458,66 +15184,67 @@ define <3 x bfloat> @v_maximumnum_v3bf16_no_ieee(<3 x bfloat> %x, <3 x bfloat> % ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v6, v6 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.h, v4.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v8 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v7, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v9, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v2.l, v0.l, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v4.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v7, v7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v7, v7 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v5.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v5.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v5.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v3, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.h, v4.h, s2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v3bf16_no_ieee: @@ -15589,77 +15316,80 @@ define <3 x bfloat> @v_maximumnum_v3bf16_no_ieee(<3 x bfloat> %x, <3 x bfloat> % ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v6, v6 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.h, v4.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v7, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v8 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v4.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v7, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v1.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s3, v9, v11 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v7, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v2.l, v0.l, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v4.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v5.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v5.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v5, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v5.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v5.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v3, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v0.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v0, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v4.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.h, v4.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v3bf16_no_ieee: @@ -16117,85 +15847,83 @@ define <4 x bfloat> @v_maximumnum_v4bf16_no_ieee(<4 x bfloat> %x, <4 x bfloat> % ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v0.h, v2.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v4.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v10, v10 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v1.h, v3.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v10, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v5.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v9, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v7.l, v6.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v8.l -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v11, v12 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v9, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v7.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v0.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v8, v8 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v2.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v7.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v5.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v1.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v8.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v8, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v8.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v8.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v3, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v7.h, s4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v5.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v5.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_maximumnum_v4bf16_no_ieee: @@ -16290,99 +16018,98 @@ define <4 x bfloat> @v_maximumnum_v4bf16_no_ieee(<4 x bfloat> %x, <4 x bfloat> % ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v0.h, v2.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v4.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v10, v10 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v4.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v10, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v1.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v5.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v6.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v9, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v1.h, v3.h, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v7.l, v6.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v8.l -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s1, v11, v12 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v7.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v9, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v1.l, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v8, v8 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v1.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v7.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v2.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e32 vcc_lo, v6, v7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0, v7.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v5.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v0.h, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v1.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v8.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v8.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s0, v8, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v8.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v8.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v3, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v7.h, s4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v0.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_gt_f32_e64 s2, v0, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0, v6.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0, v5.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v5.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v0.h, s1 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_maximumnum_v4bf16_no_ieee: diff --git a/llvm/test/CodeGen/AMDGPU/minimumnum.bf16.ll b/llvm/test/CodeGen/AMDGPU/minimumnum.bf16.ll index 678d0a4..416a601 100644 --- a/llvm/test/CodeGen/AMDGPU/minimumnum.bf16.ll +++ b/llvm/test/CodeGen/AMDGPU/minimumnum.bf16.ll @@ -120,34 +120,29 @@ define bfloat @v_minimumnum_bf16(bfloat %x, bfloat %y) { ; GFX11-TRUE16-LABEL: v_minimumnum_bf16: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v1, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_bf16: @@ -183,40 +178,34 @@ define bfloat @v_minimumnum_bf16(bfloat %x, bfloat %y) { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v2, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v1, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_bf16: @@ -344,21 +333,21 @@ define bfloat @v_minimumnum_bf16_nnan(bfloat %x, bfloat %y) { ; GFX11-TRUE16-LABEL: v_minimumnum_bf16_nnan: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v3, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v3, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.l, v0.l, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v0.l, s0 ; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_bf16_nnan: @@ -386,25 +375,25 @@ define bfloat @v_minimumnum_bf16_nnan(bfloat %x, bfloat %y) { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v3, v2 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, v0.l +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v3, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.l, v0.l, vcc_lo ; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v0.l, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v0.l, vcc_lo ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_bf16_nnan: @@ -639,58 +628,46 @@ define <2 x bfloat> @v_minimumnum_v2bf16(<2 x bfloat> %x, <2 x bfloat> %y) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v5, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v4.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v4, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.h, v4.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v4.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v1.l, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.l, v3.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v3, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v3.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v3.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v2.l, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v4, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v1.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v2bf16: @@ -747,62 +724,56 @@ define <2 x bfloat> @v_minimumnum_v2bf16(<2 x bfloat> %x, <2 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v5, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v1.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v4, v6 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v4.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v4, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.h, v4.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v1.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v1.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.l, v3.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v3, v4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v3.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v3.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v2.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v2bf16: @@ -1024,34 +995,29 @@ define <2 x bfloat> @v_minimumnum_v2bf16_nnan(<2 x bfloat> %x, <2 x bfloat> %y) ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v0 ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.h -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.h ; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v3, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v0.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, s2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v3.h, v0.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v1.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.h, v0.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.h, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.h, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v2bf16_nnan: @@ -1097,36 +1063,35 @@ define <2 x bfloat> @v_minimumnum_v2bf16_nnan(<2 x bfloat> %x, <2 x bfloat> %y) ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l ; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v0 ; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.h -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.h ; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v3, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v0.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v3.h, v0.l, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v1.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.h, v0.l, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, s0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.h, v0.h, s1 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.h, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v2bf16_nnan: @@ -1459,66 +1424,67 @@ define <3 x bfloat> @v_minimumnum_v3bf16(<3 x bfloat> %x, <3 x bfloat> %y) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v6, v6 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.h, v4.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v8 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v7, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v9, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v2.l, v0.l, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v4.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v7, v7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v7, v7 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v5.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v5.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v5.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v3, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.h, v4.h, s2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v3bf16: @@ -1590,77 +1556,80 @@ define <3 x bfloat> @v_minimumnum_v3bf16(<3 x bfloat> %x, <3 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v6, v6 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.h, v4.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v7, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v8 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v7, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v9, v11 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v7, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v2.l, v0.l, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v4.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v5.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v5.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v5.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v3, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v0.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v0, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v4.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.h, v4.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v3bf16: @@ -1957,41 +1926,40 @@ define <3 x bfloat> @v_minimumnum_v3bf16_nnan(<3 x bfloat> %x, <3 x bfloat> %y) ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v3 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v4 ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v7, v6 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v9, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.h -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v0.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v3.l, v1.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.h, v1.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v5 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.h, v1.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.l, v0.l, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v8 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.h, v0.l, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v2.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v0.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v0.h, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v3bf16_nnan: @@ -2047,48 +2015,50 @@ define <3 x bfloat> @v_minimumnum_v3bf16_nnan(<3 x bfloat> %x, <3 x bfloat> %y) ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v3 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v4 ; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v7, v6 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v3.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v9, v8 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v3.l, v1.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.l, v0.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.h -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v0.h, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.h, v1.l, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v8 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v5 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v2.h -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.h, v1.l, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.h, v0.l, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.h, v2.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v8 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v0.l, s0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v0.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v0.h, s0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v3bf16_nnan: @@ -2528,85 +2498,83 @@ define <4 x bfloat> @v_minimumnum_v4bf16(<4 x bfloat> %x, <4 x bfloat> %y) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v0.h, v2.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v4.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v10, v10 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v1.h, v3.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v10, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v5.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v9, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v7.l, v6.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v8.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v11, v12 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v9, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v7.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v0.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v8, v8 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v2.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v7.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v5.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v1.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v8.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v8, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v8.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v8.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v3, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v7.h, s4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v5.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v5.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v4bf16: @@ -2701,99 +2669,98 @@ define <4 x bfloat> @v_minimumnum_v4bf16(<4 x bfloat> %x, <4 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v0.h, v2.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v4.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v10, v10 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v10, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v1.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v5.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v6.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v9, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v1.h, v3.h, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v7.l, v6.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v8.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v11, v12 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v7.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v9, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v1.l, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v8, v8 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v1.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v7.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v2.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v7.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v5.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v0.h, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v1.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v8.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v8.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v8, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v8.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v8.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v3, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v7.h, s4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v0.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v0, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v6.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v5.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v5.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v0.h, s1 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v4bf16: @@ -3181,53 +3148,52 @@ define <4 x bfloat> @v_minimumnum_v4bf16_nnan(<4 x bfloat> %x, <4 x bfloat> %y) ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v3 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v2 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v5, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v7, v6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v1.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.l, v1.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v1.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v7, v6 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v1.h, s4 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v6, v8 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v10, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, s4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.h, v1.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.h, v1.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v2.h, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.h -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v3.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v3.h -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v1.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v3.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.h, v1.h, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v5.h, v0.l, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v2.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.h, v0.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v2.h, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v3.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v8 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.h, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v1.h, s2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.h, v0.h, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v4bf16_nnan: @@ -3296,62 +3262,63 @@ define <4 x bfloat> @v_minimumnum_v4bf16_nnan(<4 x bfloat> %x, <4 x bfloat> %y) ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v3 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.l ; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v3 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v2 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v5, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v7, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v1.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v5, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.l, v1.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v1.h, s4 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v6, v8 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.h -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v1.l, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v10, v7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v7, v6 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, s4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.h, v1.l, s0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.h, v1.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v2.h, v0.h, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.h -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v3.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v3.h -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v1.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v3.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v4 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.h, v1.h, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.h -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v3.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v5.h, v0.l, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v0.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v3.l, v2.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v8 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.h, v0.l, s0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v2.h, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.h, v0.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) ; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v4.l, v1.l, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v0.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v1.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.h, v0.h, s0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v4bf16_nnan: @@ -3984,125 +3951,120 @@ define <6 x bfloat> @v_minimumnum_v6bf16(<6 x bfloat> %x, <6 x bfloat> %y) { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v2 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v4 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v0 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v7, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v9, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v9.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v2.h, v5.h, vcc_lo ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v5 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v10, v10 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v5.h, v6.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v1.h, v4.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v5.h, v9.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v9.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v13, v13 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v9, v8 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v13, v13 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v8.h, v9.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v9.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v6.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v9 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v1.h, v4.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v4.h, v9.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v9.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v9, v10 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v10.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v10.h, v9.h, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v9.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v7.l ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v4.h, v8.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v9.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v5.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v0.h, v3.h, s2 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v13, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v14 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v8.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v6.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v9.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v7.l, v6.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v12, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v3.h, v10.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v10.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v11.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v9.l, v8.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v12.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v11.l, v6.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v13.l, v8.l, s3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.h, v7.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v9.l, s4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v2.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v8, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v15 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v14, v16 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v11.l, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v8, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v12.l, v10.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v10.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v1.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v11, v7 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v0.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v5.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v12.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v10, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v14, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v4.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v3.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v5.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v11 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 ; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v13.l, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v7.l, v3.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v0.h, v3.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v3.h, v9.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v9.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v9, v11 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v11.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v11.h, v9.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v8.l, v9.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v8.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v2.l, v5.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v5.l, v12.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v12, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v9.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v9.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v12.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v5, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.h, v9.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v4.l, s3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v9.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v12, v12 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v4.l, v1.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v1.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v2.h, v11.h, s7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v4.h, s0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v1, v9 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v9.h, v1.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v9.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v1.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v5, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.h, v10.h, s6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.h, v9.h, s3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v1.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v3.l, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v9.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.l, v1.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v3.l, v0.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v0.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v0, v9 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v9.h, v0.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v9.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v8.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v0.h, v9.h, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.h, v8.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v6.l, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s4 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v6bf16: @@ -4233,142 +4195,141 @@ define <6 x bfloat> @v_minimumnum_v6bf16(<6 x bfloat> %x, <6 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v2 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v5 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v1 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v4 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v0 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v7, v7 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v9, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v9.l +; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v2.h, v5.h, vcc_lo ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v5 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v10, v10 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v5.h, v6.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v5.h, v9.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v9.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v13, v13 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v9, v8 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v13, v13 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v1.h, v4.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v7.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v4.h, v8.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v9.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v5.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v0.h, v3.h, s2 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v13, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v14 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v8.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v6.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v9.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v8.h, v9.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v7.l, v6.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v12, v13 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v9.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v6.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v9 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v3.h, v10.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v10.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v7.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v11.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v1.h, v4.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v9.l, v8.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, v12.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v11.l, v6.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v15 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v13.l, v8.l, s3 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v13.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.h, v7.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v9.l, s4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v8, v8 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v15 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v14, v16 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v4.h, v9.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v9.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v9, v10 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v10.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v11.l, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v8, v8 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v12.l, v10.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v10.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v10.h, v9.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v9.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v10.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v1.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v11, v7 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v0.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v12.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v9.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v7.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v9 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v0.h, v3.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v5.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v12.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v10, v9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v3.h, v9.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v9.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v9, v11 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v11.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v11.h, v9.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v14, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v8 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v4.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v1.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v8.l, v9.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v8.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v2.l, v5.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v5.l, v12.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v12.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v12, v9 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v9.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v9.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v12.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v5, v5 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.h, v9.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v4.l, s3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v9.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v12, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v4.l, v1.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v1.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v2.h, v11.h, s7 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v1, v9 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v9.h, v1.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v9.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v1.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v5, v5 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.h, v10.h, s6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.h, v9.h, s3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v3.l, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v9.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v7.l, v1.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v3.l, v0.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v0.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v3.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.h, s3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v0, v9 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v1.h, v5.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.l, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v9.h, v0.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v9.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v8.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v11 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v9 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v0.h, v9.h, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.h, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.h, v8.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v6.l, v0.h, s1 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v13.l, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v7.l, v3.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s4 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v6bf16: @@ -5250,171 +5211,160 @@ define <8 x bfloat> @v_minimumnum_v8bf16(<8 x bfloat> %x, <8 x bfloat> %y) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v6 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff0000, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v3.h, v7.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v2.h, v6.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v12.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v3.h, v7.h, vcc_lo +; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v7.h, v8.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v6.h, v9.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v1.h, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v8.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v9.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v15, v17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v5.h, v12.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v13, v18 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v10.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v10.l, v8.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v11.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v11.l, v9.l, s3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v14.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v13.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v13.l, v8.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v15.l, v9.l, s4 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v7.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v17, v17 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v11 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v17, v17 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v11.h, v12.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.l, v12.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v8.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v2.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v6.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0x8000, v13.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v13.h, v12.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.l, v12.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v9.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v14, v14 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v1.h, v5.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v5.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v14 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v14.h, v12.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.l, v12.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v10.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v16, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v3 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v0.h, v4.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v4.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v15 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v15.h, v12.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v12.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v11.l, v12.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v11.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v3.l, v7.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v7.l, v16.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v16, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v12.h, v16.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.l, v16.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v7, v7 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v1 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.h, v12.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v3.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.l, v6.l, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v12.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v16, v16 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v6.l, v2.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v2.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v10.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v16, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v6.h, s0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v2, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v12.h, v2.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v12.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v2.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v7, v7 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.h, v13.h, s8 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.h, v12.h, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v5.l, s5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v12.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.l, v2.h, s2 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v19 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v8.h, v11.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v20 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v0.h, v4.h, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v16, v16 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v11, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v13.l, v8.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v15.l, v9.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.h, v10.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v14.l, v12.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v7 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v7.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v9.l, v12.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v14.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v16 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v13, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v14.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v3.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v8.l, v10.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v6.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v11.l, v10.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v12, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.h, v8.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v7.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v5 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v5.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.l, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v10 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v15, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v3.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v4.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v1.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v6.l, v2.l, s3 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.h, v7.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v14, v13 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v5.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v6.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v7.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v13 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v11.l, v3.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v4.l, v4.h, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v1.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v7.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v5.l, v1.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v5.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s5, v1, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v12.h, v1.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v12.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v1.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v7, v7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v12.h, s5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v4.l, s6 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v12.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v9 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v8 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v4.l, v0.h, s7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s6, v0, v12 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0x8000, v12.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v12.h, v0.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v11.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.h, v11.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v14.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v15.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v12.h, s9 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v10.h, v14.h, s6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v15.h, s7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v8.l, v1.h, s1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v10.l, v0.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v11.l, v5.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v4.l, s1 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v8bf16: @@ -5577,201 +5527,187 @@ define <8 x bfloat> @v_minimumnum_v8bf16(<8 x bfloat> %x, <8 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v7 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v6 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v6 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v5 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff0000, v5 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v12.l +; GFX12-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v3.h, v7.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v2.h, v6.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v3.h, v7.h, vcc_lo +; GFX12-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v4 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v7.h, v8.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v8.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v6.h, v9.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v10.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v1.h, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v8.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v9.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v15, v17 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v5.h, v12.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v13, v18 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v10.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v10.l, v8.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v11.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v11.l, v9.l, s3 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v20.l, v14.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v13.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v13.l, v8.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v15.l, v9.l, s4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v10.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v19 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v8.h, v11.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v20 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v0.h, v4.h, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v16, v16 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v11, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v7.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v17, v17 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v6 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v11 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v17, v17 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v13.l, v8.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v11.h, v12.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v15.l, v9.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.h, v10.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v14.l, v12.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v7 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v12.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v8.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v7.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, v9.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.l, v12.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v8.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v12 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v9.l, v12.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v14.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v2.h, v6.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v16 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v13, v15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v6.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0x8000, v13.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v14.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v6 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v3.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v11 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v8.l, v10.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v10.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v6.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v12, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v14 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v11.l, v10.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v8.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v12, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.h, v8.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v11.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v6.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v7.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v5 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v13.h, v12.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v5.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v13, v13 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.l, v12.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v9.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v14, v14 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v12 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v1.h, v5.h, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v5.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v14 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v14.h, v12.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.l, v12.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v10.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v16, v16 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v3 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v12 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.l, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v14, v14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v0.h, v4.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v10 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v15, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v4.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v15 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v3.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v4.l, v0.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v1.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v7.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v15.h, v12.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, v12.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v6.l, v2.l, s3 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v11.l, v12.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v11.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.h, v7.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v8.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v3.l, v7.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v7.l, v16.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v16, v12 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v12.h, v16.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.l, v16.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v7, v7 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v1 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v5 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.h, v12.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v3.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.l, v6.l, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v12.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v16, v16 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v12 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v6.l, v2.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v2.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v16, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v2, v12 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v12.h, v2.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v12.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v2.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v7, v7 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.h, v13.h, s8 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v14, v13 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v6.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v5.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.h, v12.h, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v5.l, s5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v12.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v9.l, v2.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v12 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v5.l, v1.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v1.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v6.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v2.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v5.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v5.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s5, v1, v12 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v12.h, v1.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v12.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v1.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v7, v7 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v12.h, s5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v4.l, s6 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v12.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v12 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v4.l, v0.h, s7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v0.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v7.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v5.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s6, v0, v12 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0x8000, v12.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v13 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v11.l, v3.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v12.h, v0.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v11.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v4.l, v4.h, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v1.h, s3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.h, v11.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v14.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v15.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v12.h, s9 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.h, v0.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v10.h, v14.h, s6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v15.h, s7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v8.l, v1.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v10.l, v0.h, s3 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v7.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v1, v2 :: v_dual_mov_b32 v2, v9 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v8 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v11.l, v5.l, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v4.l, s1 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v8bf16: @@ -7391,341 +7327,314 @@ define <16 x bfloat> @v_minimumnum_v16bf16(<16 x bfloat> %x, <16 x bfloat> %y) { ; GFX11-TRUE16-LABEL: v_minimumnum_v16bf16: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v16, v7 :: v_dual_mov_b32 v17, v6 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v18, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v7 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff0000, v15 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v6 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v14 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v16 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v17 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v13 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v15.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v15.h, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v19 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v17.h, v14.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v18 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v14.h, v7.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v22, v21 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v7.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v19.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v18.h, v13.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.l, v6.l, v5.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v22.l, v13.h, v20.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v21.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v21.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v24, v23 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v22.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v20.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v6.l, s1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v23.l, v19.l, v7.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v26 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v19.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v23.l, v7.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v21.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v12 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v23.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v24, v25 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v4.h, v12.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v22.l, v20.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v20.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v19.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v12.h, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v21 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v6.l, v20.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v11 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v22.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v3.h, v11.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v23.l, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v26, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v19.l, v22.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.l, v11.h, v20.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v20.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v7.l, v5.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v21.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v19.l, v5.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v13 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v12 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v20, v20 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v7.h, v15.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v21, v21 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v16.l +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.h, v15.h, v16.h, vcc_lo +; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v22, v22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v3 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v16, v17 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v16.l ; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v19.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v7.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v22, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v10.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v16.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.h, v16.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v23, v23 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v24, v24 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v22.l, v21.l, v20.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v20.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v7.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v10.h, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v23 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v22.l, v20.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v9 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v23.l, v1.h, v9.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.h, v19.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v9.h, v23.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v20.l, v21.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v27, v25 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v5.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.h, v22.l, v19.l, vcc_lo -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v7.l, v6.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v20.l, v6.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v20.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v21, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v0.h, v8.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v7.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.l, v5.l, v23.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v23.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v8.h, v19.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v16 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v24 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v21.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v21.l, v23.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v22, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.l, v16.l, v15.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v22, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v20.l, v6.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v23, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v15.l, v16.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v7.l, v19.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v16.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v22, v22 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v14 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.l, v14.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v22, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v19.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v25 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v24, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v17.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v21.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v7.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v22 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v6.l, v16.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v16.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.l, v16.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v6.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v22, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v18.l, v13.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v6.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.l, v14.l, v17.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v17.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v13.l, v15.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v24 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v16.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v16.l, v17.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v14.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v12 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v12.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.h, v14.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v18, v17 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v6.l, v15.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v11.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v5.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v15.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v18, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v3.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v21 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v5.l, v15.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v17, v17 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v6.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v17.l, v16.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v17.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff0000, v9 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v6.h, v14.h, s0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v8 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v29, v29 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v15 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v18.h, v14.h, v16.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v7 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v32, v32 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v16, v18 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v14 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v18.l, v18.h, v16.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v32, v32 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.h, v18.l, v16.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v18.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v20, v20 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v32, v32 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v5.h, v13.h, s1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v32, v32 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v11 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v25.h, v13.h, v16.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v16.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v32, v32 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v10 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v16, v25 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0x8000, v25.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v32, v32 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v25.h, v16.h, s1 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.h, v20.l, v16.h, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v20.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v21, v21 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v32, v32 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v16.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v10.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v4.h, v12.h, s2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v26.h, v12.h, v16.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v16, v26 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.l, v26.h, v16.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.h, v21.l, v16.h, s3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v21.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v22, v22 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v3.h, v11.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v27.h, v11.h, v16.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v16, v27 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v22.l, v27.h, v16.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v22.h, v22.l, v16.h, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v22.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v23, v23 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v2.h, v10.h, s4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v28.h, v10.h, v16.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v16, v28 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v23.l, v28.h, v16.h, s4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v23.h, v23.l, v16.h, s5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v23.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v24, v24 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v1.h, v9.h, s5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v29.h, v9.h, v16.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s5, v16, v29 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v24.l, v29.h, v16.h, s5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v24.h, v24.l, v16.h, s6 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v24.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v30, v30 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v0.h, v8.h, s6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.h, v8.h, v16.h, s7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s6, v16, v30 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v30.h, v16.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v25.l, v15.h, v16.h, s7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v15.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v15.l, s6 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v16.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v7.h, s8 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0x8000, v7.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s7, v7, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v16.h, v7.h, s7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v7.l, v7.h, s8 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v6.h, v16.h, s7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v7.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v14.l, s8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v14.l, v6.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0x8000, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v14.h, s7 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s8, v6, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v16.h, v6.h, s8 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v6.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v5.h, v16.h, s8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v6.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.l, v13.l, s9 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v13.l, v5.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0x8000, v5.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v13.h, s8 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s9, v5, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v5.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v5.l, v5.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v4.h, v16.h, s9 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v5.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v12.l, s10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v12.l, v4.h, s11 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0x8000, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v12.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s10, v4, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v16.h, v4.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v4.l, v4.h, s11 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v3.h, v16.h, s10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.l, v11.l, s11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v11.l, v3.h, s12 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0x8000, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v11.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s11, v3, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v16.h, v3.h, s11 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.l, v3.h, s12 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v2.h, v16.h, s11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v3.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.l, v10.l, s12 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v10.l, v2.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0x8000, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v10.h, s11 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s12, v2, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v16.h, v2.h, s12 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v2.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v31, v31 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v1.h, v16.h, s12 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v9.l, s13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v9.l, v1.h, s14 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v31, v31 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0x8000, v1.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v9.h, s12 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s13, v1, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.h, v0.l, v8.l, s14 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v31, v31 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v16.h, v1.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0x8000, v17.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.l, v1.h, s15 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v19.l, v17.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0x8000, v18.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v20.h, v25.h, s17 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0x8000, v32.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v16.h, s16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v17.l, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v19.h, v18.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v26.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v20.l, v1.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v8.l, v32.h, s14 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v18.l, v0.h, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v21.h, v26.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v28.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v27.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v32, v16 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s13 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v21.l, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v29.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v23.h, v28.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v16.h, v32.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v16.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v22.h, v27.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v24.h, v29.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v30.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v8.l, v32.h, s15 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v23.l, v1.h, s4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v22.l, v0.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v24.l, v8.h, s5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v25.l, v30.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v9.l, v16.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v8.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v15.h, v0.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v16 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v17, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v12.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v16.l, v4.h, s0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v17, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v13.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v10.l, v2.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v11.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v21, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v9.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v8.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v21, v21 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v10.l, v3.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v8.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v12.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v22, v17 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v0.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v11.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.l, v2.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v11.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v18, v17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v22, v21 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v9.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v13.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v3.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v4.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v8.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v9.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v11.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v9.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v21 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v17 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.l, v11.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v2.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v8.l, v1.h, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.l, v10.l, v0.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.l, v13.l, v2.h, s3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, v16 :: v_dual_mov_b32 v1, v14 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, v15 :: v_dual_mov_b32 v3, v20 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v19 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v8.h, vcc_lo ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v16bf16: @@ -8044,406 +7953,355 @@ define <16 x bfloat> @v_minimumnum_v16bf16(<16 x bfloat> %x, <16 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v16, v7 :: v_dual_mov_b32 v17, v6 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v18, v5 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v15 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v7 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff0000, v15 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v6 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v14 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v16 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v17 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v13 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v15.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v19.l, v5.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v15.h, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v19 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v6.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v17.h, v14.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v14.h, v7.l, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v22, v21 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v7.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v23.l, v19.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v18.h, v13.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v5.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.l, v6.l, v5.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v22 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v22.l, v13.h, v20.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v21.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v21.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v24, v23 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v26.l, v22.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v20.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v6.l, s1 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v4 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v23.l, v19.l, v7.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v26 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v19.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v23.l, v7.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v21.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v12 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v23.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v24, v25 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v4.h, v12.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v22.l, v20.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v20.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v19.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v12.h, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v21 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v6.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v6.l, v20.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v20.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v11 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v22.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v20 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v3.h, v11.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v23.l, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v26, v21 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v19.l, v22.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.l, v11.h, v20.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v20.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v7.l, v5.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v23.l, v21.l -; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v19.l, v5.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v13 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v12 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v20, v20 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v5 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v7.h, v15.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v21, v21 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v16.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.h, v15.h, v16.h, vcc_lo +; GFX12-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v4 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v22, v22 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v26.l, v16.l +; GFX12-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v3 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v16, v17 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v11 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v27.l, v16.l ; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v19.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v7.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v22, v23 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v28.l, v16.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.h, v10.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v25 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.h, v16.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v23, v23 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v24, v24 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v1 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v22.l, v21.l, v20.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v20.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v7.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v10.h, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v23 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v22.l, v20.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v9 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v26.l, v22.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v23 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v23.l, v1.h, v9.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v24, v24 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v21.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v26 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.h, v19.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v9.h, v23.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v20.l, v21.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v24 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v27, v25 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v23.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v24.l, v5.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.h, v22.l, v19.l, vcc_lo -; GFX12-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v7.l, v6.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v8 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v20.l, v6.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v20.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v21, v24 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v0.h, v8.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v7.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v25 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.l, v5.l, v23.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v23.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v8.h, v19.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v16 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v24 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v21.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v21.l, v23.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v23.l, v19.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v24.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v22, v22 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v15 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v5.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.l, v16.l, v15.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v22, v22 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v25 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v20.l, v6.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v23, v24 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v15.l, v16.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v22 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v17 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v7.l, v19.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v19.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v24.l, v16.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v22, v22 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v14 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v15.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.l, v14.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v22, v22 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v19.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v25 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v24, v23 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v17.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v21.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v7.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v22 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v17.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v14.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v18 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v6.l, v16.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v16.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v13 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.l, v16.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v6.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v22, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v18.l, v13.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v24.l, v7.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v6.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.l, v14.l, v17.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v17.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v13.l, v15.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v24 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v16.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v16.l, v17.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v14.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v12 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v17.l, v16.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v17.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l +; GFX12-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff0000, v9 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v6.h, v14.h, s0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v8 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v29, v29 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v18.h, v14.h, v16.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v7 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v32, v32 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v16, v18 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v14 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v12.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.h, v14.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v18.l, v18.h, v16.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v32, v32 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v13 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.h, v18.l, v16.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v18.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v20, v20 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v32, v32 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v16 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v3 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v18, v17 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v11 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v6.l, v15.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v5.h, v13.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v32, v32 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v11 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v25.h, v13.h, v16.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v32, v32 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v10 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v16, v25 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0x8000, v25.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v32, v32 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v11.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v13, v13 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v5.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v15.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v18, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v25.h, v16.h, s1 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v9 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.h, v20.l, v16.h, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v20.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v21, v21 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v32, v32 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v32.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v16 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v3.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v21 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v3.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v5.l, v15.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v17, v17 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v6.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v4.h, v12.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v26.h, v12.h, v16.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v16, v26 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v10.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v17, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v13 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v12.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v4.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v16.l, v4.h, s0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v17, v18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.l, v26.h, v16.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.h, v21.l, v16.h, s3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v21.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v22, v22 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v16 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v13.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v10.l, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v11.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v21, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v8 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v3.h, v11.h, s3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v27.h, v11.h, v16.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v16, v27 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v9.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v18, v18 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v8.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v21, v21 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v12.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v22.l, v27.h, v16.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v22.h, v22.l, v16.h, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v22.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v23, v23 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v16 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v1.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v10.l, v3.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v8.l, v0.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v12.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v22, v17 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v9.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v1.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v0.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v11.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v2.h, v10.h, s4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v28.h, v10.h, v16.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v16, v28 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.l, v2.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v11.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v18, v17 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v23.l, v28.h, v16.h, s4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v23.h, v23.l, v16.h, s5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v23.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v24, v24 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v16 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v22, v21 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v11 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v9.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v13.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v1.h, v9.h, s5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v29.h, v9.h, v16.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s5, v16, v29 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v3.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v4.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v8.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v9.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v24.l, v29.h, v16.h, s5 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v24.h, v24.l, v16.h, s6 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v24.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v30, v30 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v30.l, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v0.h, v8.h, s6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.h, v8.h, v16.h, s7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s6, v16, v30 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v30.h, v16.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v25.l, v15.h, v16.h, s7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v15.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v15.l, s6 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v15.l, v7.h, s8 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0x8000, v7.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s7, v7, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v16.h, v7.h, s7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v16.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v7.l, v7.h, s8 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v5 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v6.h, v16.h, s7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v7.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v14.l, s8 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v14.l, v6.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0x8000, v6.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v11.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v9.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v14.h, s7 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s8, v6, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v16.h, v6.h, s8 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v6.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v5.h, v16.h, s8 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v6.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.l, v13.l, s9 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v16 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v13.l, v5.h, s10 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0x8000, v5.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v13.h, s8 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s9, v5, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v16.h, v5.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0x8000, v16.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v5.l, v5.h, s10 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v4.h, v16.h, s9 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v5.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v12.l, s10 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v12.l, v4.h, s11 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0x8000, v4.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v18 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v21 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v17 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v12 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.l, v11.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v12.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s10, v4, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v16.h, v4.h, s10 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v4.l, v4.h, s11 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v2 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v3.h, v16.h, s10 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v3.l, v11.l, s11 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v16 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v11.l, v3.h, s12 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0x8000, v3.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v11.h, s10 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s11, v3, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v16.h, v3.h, s11 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0x8000, v16.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.l, v3.h, s12 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v1 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v2.h, v16.h, s11 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v3.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v2.l, v10.l, s12 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v10.l, v2.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0x8000, v2.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v2.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v8.l, v1.h, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.l, v10.l, v0.h, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.l, v13.l, v2.h, s3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v0, v16 :: v_dual_mov_b32 v1, v14 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v2, v15 :: v_dual_mov_b32 v3, v20 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v10.h, s11 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s12, v2, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v16.h, v2.h, s12 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v2.l, v2.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v31, v31 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v1.h, v16.h, s12 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.l, v9.l, s13 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v16.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v16 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v19 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v9.l, v1.h, s14 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v31, v31 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0x8000, v1.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v31, 16, v8 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v9.h, s12 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s13, v1, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.h, v0.l, v8.l, s14 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0x8000, v16.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v31, v31 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v16.h, v1.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0x8000, v17.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v1.l, v1.h, s15 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v19.l, v17.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0x8000, v18.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v20.h, v25.h, s17 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0x8000, v32.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v16.h, s16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v1.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v17.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v19.h, v18.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v26.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v20.l, v1.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v16 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v8.l, v32.h, s14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v18.l, v0.h, s0 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v21.h, v26.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v28.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v27.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v32, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s13 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v21.l, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v29.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v23.h, v28.h, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v16.h, v32.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v16.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v22.h, v27.h, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v24.h, v29.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v30.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v8.l, v32.h, s15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v23.l, v1.h, s4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v22.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v24.l, v8.h, s5 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v25.l, v30.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v9.l, v16.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.h, v8.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v15.h, v0.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v16 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v8.h, vcc_lo ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v16bf16: @@ -11736,666 +11594,619 @@ define <32 x bfloat> @v_minimumnum_v32bf16(<32 x bfloat> %x, <32 x bfloat> %y) { ; GFX11-TRUE16-LABEL: v_minimumnum_v32bf16: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v14 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff0000, v15 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v30 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v7 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v85, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: scratch_load_b32 v68, off, s32 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v15 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff0000, v14 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v30 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v29 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v36.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v36.l ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v33, v33 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v13 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v34, v34 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v13 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff0000, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v14.h, v30.h, s1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v11 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v10 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v9 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v8 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v24 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v6 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v5 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v69, 0xffff0000, v4 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v71, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v86, 0xffff0000, v16 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v55, v55 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s29, v85, v85 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v55.l, v30.h, v32.l, s2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v29 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v28 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff0000, v27 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v36.l +; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v12 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v28 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff0000, v11 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v27 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff0000, v26 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v25 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v23 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v68, 0xffff0000, v21 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v70, 0xffff0000, v20 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v80, 0xffff0000, v19 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v81, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v83, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v35, v35 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v37, v37 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v39, v39 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v49, v49 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v51, v51 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v53, v53 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v54, v54 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v65, v65 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v67, v67 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v69, v69 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v71, v71 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s40, v86, v86 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v54.l, v0.h, v16.h, s29 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.l, v32.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.l, v55.l -; GFX11-TRUE16-NEXT: v_and_b32_e32 v82, 0xffff0000, v18 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v84, 0xffff0000, v17 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v36, v36 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v38, v38 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v48, v48 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v50, v50 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v52, v52 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s16, v64, v64 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v66, v66 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v68, v68 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v70, v70 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v80, v80 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v81, v81 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v83, v83 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v13.h, v29.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.l, v12.h, v28.h, s5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.l, v11.h, v27.h, s7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.l, v10.h, v26.h, s9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.l, v9.h, v25.h, s11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.l, v8.h, v24.h, s13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.l, v7.h, v23.h, s15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.l, v6.h, v22.h, s17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.l, v5.h, v21.h, s19 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v50.l, v4.h, v20.h, s21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v51.l, v3.h, v19.h, s23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v85.l, v16.h, v54.l, s40 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v86 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v118, 16, v118 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v96, 16, v14 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v82, v82 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v84, v84 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v52.l, v2.h, v18.h, s25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v53.l, v1.h, v17.h, s27 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v64.l, v29.h, v33.l, s4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v65.l, v28.h, v34.l, s6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v66.l, v27.h, v35.l, s8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v67.l, v26.h, v36.l, s10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v68.l, v25.h, v37.l, s12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v69.l, v24.h, v38.l, s14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v70.l, v23.h, v39.l, s16 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v71.l, v22.h, v48.l, s18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v80.l, v21.h, v49.l, s20 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v81.l, v20.h, v50.l, s22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v82.l, v19.h, v51.l, s24 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.l, v54.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s40, v86, v118 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.l, v85.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v97, 16, v30 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v87, v87 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s41, v96, v96 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v87.l, v33.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.l, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v98.l, v36.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v101.l, v39.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v112.l, v50.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v113.l, v51.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v83.l, v18.h, v52.l, s26 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v84.l, v17.h, v53.l, s28 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v64.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.l, v65.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.l, v66.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v67.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v68.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.l, v69.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.l, v70.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v71.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.l, v80.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v81.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, v82.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v116, 16, v116 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v86 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s42, v97, v97 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v30.l, s41 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v32.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v99.l, v37.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v114.l, v52.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v115.l, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v87 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v96, 16, v96 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v98, 16, v98 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v101, 16, v101 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v112, 16, v112 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v113, 16, v113 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v83.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.l, v84.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v118, 16, v119 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v119, 16, v128 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v128, 16, v129 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v129, 16, v130 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v130, 16, v131 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v131, 16, v132 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v132, 16, v133 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v133, 16, v134 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v134, 16, v135 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v135, 16, v144 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v144, 16, v145 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s63, v116, v86 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v86.l, v55.l, v32.l, s40 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v117, 16, v13 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.l, v30.l, v14.l, s42 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0x8000, v55.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v99, 16, v99 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v114, 16, v114 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v115, 16, v115 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v145, 16, v146 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v146, 16, v147 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s42, v87, v118 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s43, v96, v119 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s45, v98, v129 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s56, v101, v132 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s59, v112, v135 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s60, v113, v144 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v86.l, v32.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.l, v86.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.l, v35.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v36.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0x8000, v39.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0x8000, v50.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0x8000, v51.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s46, v99, v130 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s61, v114, v145 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s62, v115, v146 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v96.l, v65.l, v34.l, s43 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v98.l, v67.l, v36.l, s45 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v101.l, v70.l, v39.l, s56 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v112.l, v81.l, v50.l, s59 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v113.l, v82.l, v51.l, s60 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v13.h, v55.l, s16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v118 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v33.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v37.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v102.l, v48.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0x8000, v52.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s14, 0x8000, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v97, 16, v97 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0x8000, v70.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0x8000, v81.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v87.l, v64.l, v33.l, s42 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v99.l, v68.l, v37.l, s46 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v114.l, v83.l, v52.l, s61 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v115.l, v84.l, v53.l, s62 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.h, v96.l, v34.l, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.h, v98.l, v36.l, s5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.l, v101.l, v39.l, s8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v101.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.h, v112.l, v50.l, s11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v112.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.l, v113.l, v51.l, s12 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v113.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v55 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v100.l, v38.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v103.l, v49.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0x8000, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v102, 16, v102 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s44, v97, v128 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v116.l, v85.l, v54.l, s63 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v87.l, v33.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v99.l, v37.l, s6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v114.l, v52.l, s13 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v114.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.l, v115.l, v53.l, s14 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v115.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.l, v34.l, v70.l, s23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v70, 16, v39 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.l, v35.h, v81.l, s26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v35.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v100, 16, v100 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v103, 16, v103 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s57, v102, v133 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v97.l, v66.l, v35.l, s44 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.h, v116.l, v54.l, s15 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v116.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v51 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0x8000, v48.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0x8000, v65.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0x8000, v66.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s47, v100, v131 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s58, v103, v134 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v102.l, v71.l, v48.l, s57 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.l, v96.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v97.l, v35.l, s4 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.l, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v52 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s14, 0, v53 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v38.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0x8000, v49.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0x8000, v64.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0x8000, v71.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v100.l, v69.l, v38.l, s47 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v103.l, v80.l, v49.l, s58 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v87.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.h, v102.l, v48.l, s9 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v102.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.h, v30.h, v65.l, s18 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v128 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v32.l, v66.l, s19 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 16, v129 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0x8000, v80.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s41, 0x8000, v85.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.h, v100.l, v38.l, s7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.l, v103.l, v49.l, s10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v103.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.l, v14.h, v64.l, s17 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 16, v119 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.h, v34.h, v71.l, s24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v71, 16, v48 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v65 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v66 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v86.l, v13.h, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.l, v35.l, v80.l, s25 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 16, v49 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.l, v37.h, v85.l, s41 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v64 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v71 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.h, v96.l, v30.h, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.h, v97.l, v32.l, s4 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0x8000, v67.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0x8000, v68.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v98.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v99.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v87.l, v38.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.h, v102.l, v38.h, s9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.h, v32.h, v67.l, s20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v67, 16, v130 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v33.l, v68.l, s21 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v68, 16, v131 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s29, 0x8000, v84.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v70 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v67 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s27, 0x8000, v82.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v68 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.l, v37.l, v84.l, s29 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.h, v101.l, v34.l, s8 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0x8000, v83.l -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v80 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.l, v36.l, v82.l, s27 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.h, v115.l, v37.l, s14 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0x8000, v69.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.l, v36.h, v83.l, s28 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v103.l, v35.l, s10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.h, v113.l, v36.l, s12 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.l, v100.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.h, v33.h, v69.l, s22 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v81 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.h, v114.l, v48.l, s13 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 16, v132 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v69 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.h, v100.l, v33.h, s7 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v25 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v39, v39 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v49, v49 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v10 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v50, v50 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff0000, v9 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v51, v51 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v8 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v24 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v23 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v22 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v21 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v20 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v52, v52 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v7 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v53, v53 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v6 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v54, v54 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v55, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v4 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v64, v64 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v3 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v19 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v18 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v17 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v81, 0xffff0000, v16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v65, v65 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v66, v66 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v67, v67 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v15 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v83, 16, v30 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v84, 16, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v85, 16, v17 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v16 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v83, v83 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s42, v86, v86 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v96.h, v0.l, v16.l, s42 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s43, 0x8000, v96.h ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v31 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v50.l, v15.h, v31.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v31.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v55, v55 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v68 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v15.h, v68.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.h, v68.h, v36.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v36, v35 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s42, 0x8000, v35.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.l, v35.h, v36.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.h, v31.l, v36.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v31.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v32, v32 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v14.h, v30.h, s0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.h, v30.h, v36.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v36, v37 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s44, 0x8000, v37.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v37.h, v36.h, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.h, v32.l, v36.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v32.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v33, v33 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v13.h, v29.h, s1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.h, v29.h, v36.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v36, v38 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v38.h, v36.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.h, v33.l, v36.h, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v33.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v34, v34 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v12.h, v28.h, s2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.h, v28.h, v36.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v36, v39 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.l, v39.h, v36.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.h, v34.l, v36.h, s3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v34.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v48, v48 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v36.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v50.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v15.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v51.l, v31.h, v50.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v54 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v31 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.h, v98.l, v32.h, s5 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v51.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.h, v99.l, v33.l, s6 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v54, v54 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.l, v31.l, v15.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v50.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v52, v53 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v31.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v11.h, v27.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.h, v27.h, v36.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v36, v48 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.l, v48.h, v36.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.l, v35.l, v36.h, s4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v35.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v49, v49 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v36.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v51.l, v50.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v52 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v51.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v32.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v32.l, v50.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v10.h, v26.h, s4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.h, v26.h, v36.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v36, v49 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.l, v49.h, v36.h, s4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.l, v38.l, v36.h, s5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v38.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v50, v50 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v9.h, v25.h, s5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v50.h, v25.h, v36.h, s6 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s5, v36, v50 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.l, v50.h, v36.h, s5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.l, v48.l, v36.h, s6 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v48.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v51, v51 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v8.h, v24.h, s6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v51.h, v24.h, v36.h, s7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s6, v36, v51 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v50.l, v51.h, v36.h, s6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v51.l, v50.l, v36.h, s7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v50.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v52, v52 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v7.h, v23.h, s7 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v52.h, v23.h, v36.h, s8 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s7, v36, v52 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v52.l, v52.h, v36.h, s7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v69.l, v52.l, v36.h, s8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v52.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v53, v53 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v6.h, v22.h, s8 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v53.h, v22.h, v36.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s8, v36, v53 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v53.l, v53.h, v36.h, s8 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v69.h, v53.l, v36.h, s9 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v53.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v54, v54 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v5.h, v21.h, s9 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v54.h, v21.h, v36.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s9, v36, v54 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v54.l, v54.h, v36.h, s9 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v70.l, v54.l, v36.h, s10 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v54.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v55, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v4.h, v20.h, s10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v55.h, v20.h, v36.h, s11 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s10, v36, v55 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v55.l, v55.h, v36.h, s10 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v70.h, v55.l, v36.h, s11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v55.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v64, v64 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v3.h, v19.h, s11 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v64.h, v19.h, v36.h, s12 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s11, v36, v64 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v64.l, v64.h, v36.h, s11 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v71.l, v64.l, v36.h, s12 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v64.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v65, v65 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v2.h, v18.h, s12 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v65.h, v18.h, v36.h, s13 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s12, v36, v65 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v65.l, v65.h, v36.h, s12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v71.h, v65.l, v36.h, s13 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v65.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v66, v66 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v1.h, v17.h, s13 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v66.h, v17.h, v36.h, s14 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s14, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s13, v36, v66 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v66.l, v66.h, v36.h, s13 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v80.l, v66.l, v36.h, s14 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v66.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v67, v67 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v0.h, v16.h, s14 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v67.h, v16.h, v36.h, s15 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v68 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s14, v36, v67 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v82.h, v15.l, v68.l, s15 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v14 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v67.l, v67.h, v36.h, s14 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v67.l, v36.h, s16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v67.l +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0x8000, v82.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s14, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v68.l, v82.h, s15 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s15, v82, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v68.l, v36.h, v82.h, s15 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v68.l, v82.h, s16 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s16, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v13 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v29 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.h, v14.h, v36.h, s15 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v68.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v14.l, v30.l, s16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v28 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s15, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v30.l, v14.h, s17 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0x8000, v14.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v38, v53 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.h, v112.l, v39.l, s11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v52 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v15.h, v51.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v31.l, v15.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v29 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v50 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v30.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v33.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v33.l, v15.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v32.l, v15.h, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v117, v117 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v31.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v29.h, v116.l, v49.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v27 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s16, v14, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v82, v82 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v36.h, v14.h, s16 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v26 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v14.l, v14.h, s17 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v12 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v25 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v29.h, v13.h, v36.h, s16 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v14.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v13.l, v29.l, s17 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v24 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s16, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v29.l, v13.h, s18 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0x8000, v13.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v23 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s17, v13, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v29.h, s16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v82, v82 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v36.h, v13.h, s17 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v22 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v13.l, v13.h, s18 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v11 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v21 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v28.h, v12.h, v36.h, s17 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v13.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v12.l, v28.l, s18 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v20 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s17, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v28.l, v12.h, s19 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0x8000, v12.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v82, v82 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v19 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s18, v12, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v13.l, v28.h, s17 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v82, v82 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v36.h, v12.h, s18 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.l, v36.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v12.l, v12.h, s19 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v10 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v27.h, v11.h, v36.h, s18 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v12.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v11.l, v27.l, s19 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s18, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v27.l, v11.h, s20 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0x8000, v11.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v27.h, s18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s19, v11, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v36.h, v11.h, s19 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0x8000, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v11.l, v11.h, s20 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v9 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v26.h, v10.h, v36.h, s19 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v11.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.l, v26.l, s20 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s19, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v26.l, v10.h, s21 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0x8000, v10.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v26.h, s19 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s20, v10, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v36.h, v10.h, s20 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0x8000, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v10.l, v10.h, s21 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v8 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v25.h, v9.h, v36.h, s20 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v10.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.l, v25.l, s21 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s20, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v25.l, v9.h, s22 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0x8000, v9.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v10.l, v25.h, s20 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s21, v9, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v36.h, v9.h, s21 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0x8000, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v9.l, v9.h, s22 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v24.h, v8.h, v36.h, s21 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v9.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.l, v24.l, s22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s21, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v24.l, v8.h, s23 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0x8000, v8.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v24.h, s21 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s22, v8, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v36.h, v8.h, s22 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0x8000, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v8.l, v8.h, s23 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v23.h, v7.h, v36.h, s22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v8.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v23.l, s23 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s22, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v23.l, v7.h, s24 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0x8000, v7.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v23.h, s22 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s23, v7, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v36.h, v7.h, s23 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0x8000, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v7.l, v7.h, s24 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v22.h, v6.h, v36.h, s23 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v7.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v22.l, s24 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s23, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v22.l, v6.h, s25 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0x8000, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v22.h, s23 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s24, v6, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v36.h, v6.h, s24 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0x8000, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v6.h, s25 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v21.h, v5.h, v36.h, s24 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.l, v21.l, s25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s24, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v21.l, v5.h, s26 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0x8000, v5.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v21.h, s24 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s25, v5, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v36.h, v5.h, s25 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0x8000, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v5.l, v5.h, s26 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v20.h, v4.h, v36.h, s25 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v5.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v20.l, s26 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s25, 0, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v20.l, v4.h, s27 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s27, 0x8000, v4.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v20.h, s25 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s26, v4, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v36.h, v4.h, s26 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0x8000, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v4.l, v4.h, s27 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v81, v81 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v18 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v18.h, v3.h, v36.h, s26 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v4.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v83.h, v3.l, v19.l, s27 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s29, v81, v81 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s26, 0, v36 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v13.l, v29.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v52, v52 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v53 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v51, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v31.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v29.l, v29.l, v13.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v52 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.l, v30.l, v14.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v13.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v29.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v28 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v31.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v28.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v31.l, v14.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v53 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v52, v51 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v28.l, v28.l, v12.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v12.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v12.h, v30.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v50 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.l, v29.l, v13.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v33.l, v15.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v13.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.l, v31.l, v12.h, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v27 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v30.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v30.l, v13.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v51, v50 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v27.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v53 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v29.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.l, v28.l, v12.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v27.l, v27.l, v11.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v11.h, v29.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v31.l, v12.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v11.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v26 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v28.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v31.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v10.l, v26.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v52 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.l, v30.l, v11.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v51, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.h, v28.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v26.l, v10.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v50 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v26.l, v27.l, v11.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v28, v28 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v25.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v28, v28 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v26.l, v11.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v52 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v51, v50 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v25.l, v9.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v37.l, v31.l, v10.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.h, v27.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v28 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v25.l, v12.l, v10.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v9.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v30.l, v26.l, v9.h, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v27 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v28 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v25.l, v10.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v50, v50 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v24 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v27, v26 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v12.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v24.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v28 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v11.l, v9.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v9.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.h, v12.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.l, v24.l, v8.l, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v7 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v26 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v10.l, v9.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v12.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v24, v24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v23 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v26 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v23.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v24, v24 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v27 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v31.l, v25.l, v8.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v26, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v23.l, v7.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.h, v11.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.l, v12.l, v8.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v22 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v22.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v11.l, v8.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v12.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v26 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v25, v24 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v22.l, v6.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v32.l, v10.l, v7.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.h, v12.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v23 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.l, v9.l, v7.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v6.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v33.l, v11.l, v6.h, vcc_lo -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v10.l, v7.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v21 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v10.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v12, v11 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v9.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v21.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v22 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v8.l, v6.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v9.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v21.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v7.l, v6.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v7.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v20 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v20.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v21 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v34.l, v10.l, v5.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v20.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.h, v8.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v9.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v19 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v19.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v8.l, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v20 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v12, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v19.l, v3.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v9.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v18 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.l, v6.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v18.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v9.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v6.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v35.l, v7.l, v4.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v18.l, v2.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v11, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.l, v8.l, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v6.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v17 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v16 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v17.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v16.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v17.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v3.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v16.l, v0.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v12, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v3.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.l, v2.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v2.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v17, v16 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v7.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v4.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v8.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v7.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v16 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v11 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v29.l, v5.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v49.l, v2.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v48.l, v8.l, v1.h, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v39.l, v6.l, v0.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v38.l, v9.l, v2.h, s3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, v29 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, v49 :: v_dual_mov_b32 v2, v48 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v39 :: v_dual_mov_b32 v4, v38 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v36 :: v_dual_mov_b32 v6, v35 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v7, v34 :: v_dual_mov_b32 v8, v33 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v32 :: v_dual_mov_b32 v10, v31 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v11, v30 :: v_dual_mov_b32 v12, v37 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v19.l, v83.h, s28 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v84, v84 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s40, v3, v3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s41, 0x8000, v83.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v36.l +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s27, v83, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v82.h, v2.l, v18.l, s28 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v19.h, v1.l, v17.l, s40 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v32.h, v37.h, s44 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v36.h, v83.h, s27 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s40, 0x8000, v82.h +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v85, v85 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v18.h, s26 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v14.h, v32.l, v1.l, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v83.h, s41 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s41, v87, v87 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.h, v36.h, s28 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v3.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v31.h, v35.h, s42 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s42, 0x8000, v39.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0x8000, v19.h +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s45, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v18.l, v82.h, s29 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s29, 0x8000, v38.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.h, v31.l, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v34.h, v39.h, s42 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v48.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s44, v82, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v33.h, v38.h, s29 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v12.h, v34.l, v1.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v37.l, v48.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v36.h, v82.h, s44 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v13.h, v33.l, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v49.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v50.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v11.h, v35.l, v1.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v82.h, s40 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v51.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v54.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v49.l, v50.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.l, s45 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v36.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v39.l, v49.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v52.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v51.l, v51.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v53.h +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v17.l, v19.h, s27 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v10.h, v38.l, v0.h, s4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v69.l, v52.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v55.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v9.h, v48.l, v1.h, s5 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v19, v36 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v36.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v69.h, v53.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v70.h, v55.h, s4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v66.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.h, v36.h, v19.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v65.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v50.l, v2.h, s6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v70.l, v54.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v52.l, v0.h, s7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.l, v16.h, v19.h, s28 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v53.l, v1.h, s8 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v71.h, v65.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v54.l, v2.h, s9 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v67.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.l, v36.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v16.h +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v64.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v65.l, v1.h, s12 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v67.h, s3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v55.l, v3.h, s10 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v36.h, v16.l, v96.h, s41 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v71.l, v64.h, s0 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.l, v80.l, v66.h, s2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v2.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v16.h, v17.l, s4 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v96, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v64.l, v0.h, s11 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v66.l, v16.l, s13 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v67.l, v15.l, s14 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v15.l, v68.l, v30.h, s15 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v16.l, v36.h, v96.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v36.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.h, v16.l, v96.h, s43 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v17.h, v17.h, v36.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v16.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v36 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v16.l, v17.h, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v32bf16: @@ -13011,753 +12822,697 @@ define <32 x bfloat> @v_minimumnum_v32bf16(<32 x bfloat> %x, <32 x bfloat> %y) { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: scratch_load_b32 v31, off, s32 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v14 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff0000, v15 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v30 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v7 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v85, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: scratch_load_b32 v68, off, s32 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v15 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff0000, v14 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v30 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v29 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v35.l, v36.l +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v37.l, v36.l ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v33, v33 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v13 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v34, v34 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v13 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff0000, v12 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v14.h, v30.h, s1 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v11 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v10 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v9 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v8 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v24 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v6 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v5 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v69, 0xffff0000, v4 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v71, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v86, 0xffff0000, v16 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v55, v55 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s29, v85, v85 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v55.l, v30.h, v32.l, s2 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v29 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v28 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff0000, v27 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v38.l, v36.l +; GFX12-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v12 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v28 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff0000, v11 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v27 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff0000, v26 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v25 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v23 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v22 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v68, 0xffff0000, v21 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v70, 0xffff0000, v20 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v80, 0xffff0000, v19 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v81, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v83, 0xffff0000, v1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v35, v35 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v37, v37 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v39, v39 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v49, v49 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v51, v51 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v53, v53 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v54, v54 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v65, v65 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v67, v67 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v69, v69 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v71, v71 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s40, v86, v86 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v54.l, v0.h, v16.h, s29 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v86.l, v32.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v118.l, v55.l -; GFX12-TRUE16-NEXT: v_and_b32_e32 v82, 0xffff0000, v18 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v84, 0xffff0000, v17 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v36, v36 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v38, v38 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v48, v48 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v50, v50 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v52, v52 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s16, v64, v64 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v66, v66 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v68, v68 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v70, v70 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v80, v80 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v81, v81 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v83, v83 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v13.h, v29.h, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.l, v12.h, v28.h, s5 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.l, v11.h, v27.h, s7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.l, v10.h, v26.h, s9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.l, v9.h, v25.h, s11 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.l, v8.h, v24.h, s13 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.l, v7.h, v23.h, s15 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.l, v6.h, v22.h, s17 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.l, v5.h, v21.h, s19 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v50.l, v4.h, v20.h, s21 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v51.l, v3.h, v19.h, s23 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v85.l, v16.h, v54.l, s40 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v86 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v118, 16, v118 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v15 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v96, 16, v14 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v82, v82 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v84, v84 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v52.l, v2.h, v18.h, s25 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v53.l, v1.h, v17.h, s27 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v64.l, v29.h, v33.l, s4 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v65.l, v28.h, v34.l, s6 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v66.l, v27.h, v35.l, s8 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v67.l, v26.h, v36.l, s10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v68.l, v25.h, v37.l, s12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v69.l, v24.h, v38.l, s14 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v70.l, v23.h, v39.l, s16 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v71.l, v22.h, v48.l, s18 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v80.l, v21.h, v49.l, s20 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v81.l, v20.h, v50.l, s22 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v82.l, v19.h, v51.l, s24 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v116.l, v54.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s40, v86, v118 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v86.l, v85.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v97, 16, v30 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v87, v87 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s41, v96, v96 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v87.l, v33.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v96.l, v34.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v98.l, v36.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v101.l, v39.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v112.l, v50.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v113.l, v51.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v83.l, v18.h, v52.l, s26 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v84.l, v17.h, v53.l, s28 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v119.l, v64.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v128.l, v65.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v129.l, v66.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v130.l, v67.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v131.l, v68.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v132.l, v69.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v133.l, v70.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v134.l, v71.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v135.l, v80.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v144.l, v81.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v145.l, v82.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v116, 16, v116 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v86 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s42, v97, v97 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v30.l, s41 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v32.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v99.l, v37.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v114.l, v52.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v115.l, v53.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v87 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v96, 16, v96 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v98, 16, v98 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v101, 16, v101 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v112, 16, v112 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v113, 16, v113 -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v146.l, v83.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v147.l, v84.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v118, 16, v119 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v119, 16, v128 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v128, 16, v129 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v129, 16, v130 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v130, 16, v131 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v131, 16, v132 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v132, 16, v133 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v133, 16, v134 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v134, 16, v135 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v135, 16, v144 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v144, 16, v145 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s63, v116, v86 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v86.l, v55.l, v32.l, s40 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v117, 16, v13 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.l, v30.l, v14.l, s42 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0x8000, v55.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v99, 16, v99 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v114, 16, v114 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v115, 16, v115 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v145, 16, v146 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v146, 16, v147 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s42, v87, v118 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s43, v96, v119 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s45, v98, v129 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s56, v101, v132 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s59, v112, v135 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s60, v113, v144 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v86.l, v32.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v118.l, v86.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v34.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v97.l, v35.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v36.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0x8000, v39.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0x8000, v50.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0x8000, v51.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s46, v99, v130 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s61, v114, v145 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s62, v115, v146 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v96.l, v65.l, v34.l, s43 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v98.l, v67.l, v36.l, s45 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v101.l, v70.l, v39.l, s56 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v112.l, v81.l, v50.l, s59 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v113.l, v82.l, v51.l, s60 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v13.h, v55.l, s16 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v55, 16, v118 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v33.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v37.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v102.l, v48.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0x8000, v52.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s14, 0x8000, v53.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v97, 16, v97 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0x8000, v70.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0x8000, v81.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v87.l, v64.l, v33.l, s42 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v99.l, v68.l, v37.l, s46 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v114.l, v83.l, v52.l, s61 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v115.l, v84.l, v53.l, s62 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.h, v96.l, v34.l, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.h, v98.l, v36.l, s5 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.l, v101.l, v39.l, s8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v39.l, v101.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.h, v112.l, v50.l, s11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v50.l, v112.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.l, v113.l, v51.l, s12 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v113.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v55 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v100.l, v38.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v103.l, v49.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0x8000, v54.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v102, 16, v102 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s44, v97, v128 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v116.l, v85.l, v54.l, s63 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v87.l, v33.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v99.l, v37.l, s6 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v114.l, v52.l, s13 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v114.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.l, v115.l, v53.l, s14 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v115.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.l, v34.l, v70.l, s23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v70, 16, v39 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.l, v35.h, v81.l, s26 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v35.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v100, 16, v100 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v103, 16, v103 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s57, v102, v133 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v97.l, v66.l, v35.l, s44 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.h, v116.l, v54.l, s15 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v54.l, v116.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v51 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0x8000, v48.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0x8000, v65.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0x8000, v66.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s47, v100, v131 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s58, v103, v134 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v102.l, v71.l, v48.l, s57 -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v128.l, v96.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v97.l, v35.l, s4 -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v129.l, v97.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v54 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v52 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s14, 0, v53 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v38.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0x8000, v49.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0x8000, v64.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0x8000, v71.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v100.l, v69.l, v38.l, s47 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v103.l, v80.l, v49.l, s58 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v119.l, v87.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.h, v102.l, v48.l, s9 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v48.l, v102.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.h, v30.h, v65.l, s18 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v65, 16, v128 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v32.l, v66.l, s19 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v66, 16, v129 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0x8000, v80.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s41, 0x8000, v85.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.h, v100.l, v38.l, s7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.l, v103.l, v49.l, s10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v49.l, v103.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.l, v14.h, v64.l, s17 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v64, 16, v119 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.h, v34.h, v71.l, s24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v71, 16, v48 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v65 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v66 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v86.l, v13.h, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.l, v35.l, v80.l, s25 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v80, 16, v49 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.l, v37.h, v85.l, s41 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v64 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v71 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.h, v96.l, v30.h, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.h, v97.l, v32.l, s4 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0x8000, v67.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0x8000, v68.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v130.l, v98.l -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v131.l, v99.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v87.l, v38.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.h, v102.l, v38.h, s9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.h, v32.h, v67.l, s20 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v67, 16, v130 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v33.l, v68.l, s21 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v68, 16, v131 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s29, 0x8000, v84.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v70 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v67 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s27, 0x8000, v82.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v68 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.l, v37.l, v84.l, s29 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.h, v101.l, v34.l, s8 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0x8000, v83.l -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v80 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.l, v36.l, v82.l, s27 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.h, v115.l, v37.l, s14 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0x8000, v69.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.l, v36.h, v83.l, s28 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v103.l, v35.l, s10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.h, v113.l, v36.l, s12 -; GFX12-TRUE16-NEXT: v_mov_b16_e64 v132.l, v100.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.h, v33.h, v69.l, s22 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v81 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.h, v114.l, v48.l, s13 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v69, 16, v132 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v69 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.h, v100.l, v33.h, s7 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v25 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v39, v39 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v39.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v49, v49 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v10 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v50, v50 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff0000, v9 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v51, v51 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v8 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v24 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v23 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v22 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v21 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v20 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v52, v52 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff0000, v7 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v53, v53 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff0000, v6 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v54, v54 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff0000, v5 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v55, v55 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v4 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v64, v64 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v64, 0xffff0000, v3 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v19 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v18 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v17 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v81, 0xffff0000, v16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v82.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v65, v65 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v65, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v66, v66 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v67, v67 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v67, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v15 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v83, 16, v30 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v84, 16, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v86, 16, v0 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v85, 16, v17 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v87, 16, v16 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v83, v83 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v83.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s42, v86, v86 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v96.l, v36.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v96.h, v0.l, v16.l, s42 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s43, 0x8000, v96.h ; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff0000, v31 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v68 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v50.l, v15.h, v31.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v31.l, s0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v55, v55 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v50.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v15.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v15.h, v68.h, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v36.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v51.l, v31.h, v50.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v54 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v54, 16, v31 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.h, v98.l, v32.h, s5 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v51.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.h, v99.l, v33.l, s6 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v54, v54 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.l, v31.l, v15.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v50.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v52, v53 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v31.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v51.l, v50.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v38, 16, v52 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v53, 16, v53 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v51.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v32.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v32.l, v50.l, s0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v38, v53 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.h, v112.l, v39.l, s11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v52 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v15.h, v51.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v31.l, v15.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v14.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v29 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v50 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v50.l, v30.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v33.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v33.l, v15.l, s2 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v32.l, v15.h, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v117, v117 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v31.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.h, v68.h, v36.h, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v36, v35 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s42, 0x8000, v35.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v29.h, v116.l, v49.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.l, v35.h, v36.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v13.l, v29.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v52, v52 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v53 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v51, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v31.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v29.l, v29.l, v13.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v52 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.l, v30.l, v14.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v13.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v29.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v28 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v14.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v31.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v52 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.h, v31.l, v36.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v31.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v32, v32 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v14.h, v30.h, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.h, v30.h, v36.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v36, v37 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s44, 0x8000, v37.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v37.h, v36.h, s0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.h, v32.l, v36.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v32.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v33, v33 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v28.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v31.l, v14.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v30.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v53 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v52, v51 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v13.h, v29.h, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.h, v29.h, v36.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v36, v38 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v28.l, v28.l, v12.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v12.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v12.h, v30.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v50 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.l, v29.l, v13.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v50.l, v28.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v11 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v33.l, v15.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v13.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v31.l, v12.h, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v52, 16, v27 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v30.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v38.h, v36.h, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.h, v33.l, v36.h, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v33.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v34, v34 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v30.l, v13.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v51, v50 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v27.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v52, v52 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v53 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v29.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v12.h, v28.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.h, v28.h, v36.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v36, v39 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.l, v28.l, v12.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v12.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v27.l, v27.l, v11.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v11.h, v29.l, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.l, v39.h, v36.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.h, v34.l, v36.h, s3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v34.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v48, v48 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v48.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v31.l, v12.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v27.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v11.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v26 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v28.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v31.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v11.h, v27.h, s3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.h, v27.h, v36.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v36, v48 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v10.l, v26.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v50, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v52 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v30.l, v11.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v51, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.h, v28.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.l, v48.h, v36.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.l, v35.l, v36.h, s4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v35.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v49, v49 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v49.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v26.l, v10.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v50 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v26.l, v27.l, v11.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v10.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v28, v28 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v25 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v11.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v26.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v51, 16, v51 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v10.h, v26.h, s4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.h, v26.h, v36.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s5, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s4, v36, v49 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v25.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v28, v28 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v26.l, v11.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v27.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v52 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v51, v50 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.l, v49.h, v36.h, s4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.l, v38.l, v36.h, s5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v38.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s5, v50, v50 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v50.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v25.l, v9.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v37.l, v31.l, v10.h, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.h, v27.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v28 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v25.l, v12.l, v10.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v27.l, v11.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v28.l, v9.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v10.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v50, 16, v8 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.l, v26.l, v9.h, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v27 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v28 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v9.h, v25.h, s5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v50.h, v25.h, v36.h, s6 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s6, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s5, v36, v50 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v25.l, v10.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v50, v50 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v24 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v27, v26 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v12.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v24.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v28 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.l, v50.h, v36.h, s5 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.l, v48.l, v36.h, s6 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v48.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s6, v51, v51 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v51.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s5, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v11.l, v9.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v9.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.h, v12.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v24.l, v8.l, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v7 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v26 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v8.h, v24.h, s6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v51.h, v24.h, v36.h, s7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s7, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s6, v36, v51 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v10.l, v9.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v26.l, v8.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v12.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v24, v24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v23 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v27.l, v10.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v26 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v50.l, v51.h, v36.h, s6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v51.l, v50.l, v36.h, s7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v50.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s7, v52, v52 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v52.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s6, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v23.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v24, v24 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v11.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v27 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v31.l, v25.l, v8.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v26, v9 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v7.h, v23.h, s7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v52.h, v23.h, v36.h, s8 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s8, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s7, v36, v52 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v23.l, v7.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v6 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.h, v11.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v24 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v12.l, v8.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v24.l, v9.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v25.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v22 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v8.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v26.l, v11.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v52.l, v52.h, v36.h, s7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v69.l, v52.l, v36.h, s8 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v52.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s8, v53, v53 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v53.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s7, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v22.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v23, v23 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v11.l, v8.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v12.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v26 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v25, v24 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v6.h, v22.h, s8 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v53.h, v22.h, v36.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s9, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s8, v36, v53 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v22.l, v6.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v32.l, v10.l, v7.h, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.h, v12.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v23 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v9.l, v7.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v5 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v33.l, v11.l, v6.h, vcc_lo -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v22 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v10.l, v7.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v21 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v22.l, v10.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v12, v11 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v9.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v21.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v22 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v53.l, v53.h, v36.h, s8 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v69.h, v53.l, v36.h, s9 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v53.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s9, v54, v54 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v54.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s8, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v8.l, v6.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.h, v9.l, s1 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v21.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v5.h, v21.h, s9 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v54.h, v21.h, v36.h, s10 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s10, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s9, v36, v54 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v7.l, v6.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v21.l, v7.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v20 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v8.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v54.l, v54.h, v36.h, s9 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v70.l, v54.l, v36.h, s10 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v54.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s10, v55, v55 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v55.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s9, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v20.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v21 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v34.l, v10.l, v5.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v4.h, v20.h, s10 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v55.h, v20.h, v36.h, s11 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s11, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s10, v36, v55 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v20.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.h, v8.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v11 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v9.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v4.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v19 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v20.l, v8.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v55.l, v55.h, v36.h, s10 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v70.h, v55.l, v36.h, s11 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v55.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s11, v64, v64 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v64.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s10, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v19.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v10, v10 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v8.l, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v9.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v20 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v12, v11 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v3.h, v19.h, s11 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v64.h, v19.h, v36.h, s12 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s12, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s11, v36, v64 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v19.l, v3.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v9.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v18 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v6.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v4.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v64.l, v64.h, v36.h, s11 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v71.l, v64.l, v36.h, s12 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v64.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s12, v65, v65 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v65.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s11, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v18.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v9.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v6.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v35.l, v7.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v2.h, v18.h, s12 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v65.h, v18.h, v36.h, s13 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s13, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s12, v36, v65 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v18.l, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v11, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.l, v8.l, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v3.h, v6.l, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v65.l, v65.h, v36.h, s12 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v71.h, v65.l, v36.h, s13 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v65.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s13, v66, v66 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v66.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s12, 0, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v17 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v16 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v9.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v1.h, v17.h, s13 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v66.h, v17.h, v36.h, s14 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s14, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s13, v36, v66 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v17.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v10, v10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v16.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v11, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v66.l, v66.h, v36.h, s13 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v80.l, v66.l, v36.h, s14 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v66.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s14, v67, v67 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v67.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s13, 0, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v0.h, v16.h, s14 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v67.h, v16.h, v36.h, s15 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v68 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s14, v36, v67 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v82.h, v15.l, v68.l, s15 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s15, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v67.l, v67.h, v36.h, s14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v67.l, v36.h, s16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v67.l +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0x8000, v82.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s14, 0, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v68.l, v82.h, s15 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s15, v82, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v68.l, v36.h, v82.h, s15 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s15, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v68.l, v82.h, s16 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s16, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v13 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v29 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v30.h, v14.h, v36.h, s15 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v68.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v14.l, v30.l, s16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v28 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s15, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v30.l, v14.h, s17 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0x8000, v14.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v27 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s16, v14, v36 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v82, v82 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v36.h, v14.h, s16 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s16, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v26 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v14.l, v14.h, s17 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s17, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v12 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v25 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v29.h, v13.h, v36.h, s16 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v14.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v13.l, v29.l, s17 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v24 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s16, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v29.l, v13.h, s18 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0x8000, v13.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v23 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s17, v13, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v17.l, v1.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v3.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v16.l, v0.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v12, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v1.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v7.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v17.l, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, v3.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v5.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.l, v14.l, v29.h, s16 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v82, v82 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v36.h, v13.h, s17 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s17, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v22 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v13.l, v13.h, s18 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s18, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v11 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v21 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v28.h, v12.h, v36.h, s17 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v13.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v12.l, v28.l, s18 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v20 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s17, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v28.l, v12.h, s19 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0x8000, v12.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v82, v82 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v82, 16, v19 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s18, v12, v36 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v4.l, v2.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v17 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v12, v11 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.l, v13.l, v28.h, s17 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v82, v82 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v36.h, v12.h, s18 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s18, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v82.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v12.l, v12.h, s19 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s19, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v10 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v27.h, v11.h, v36.h, s18 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v12.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v11.l, v27.l, s19 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s18, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v27.l, v11.h, s20 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0x8000, v11.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v2.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v17, v16 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v5 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v7.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.l, v12.l, v27.h, s18 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s19, v11, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v36.h, v11.h, s19 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s19, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v11.l, v11.h, s20 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s20, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v9 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v26.h, v10.h, v36.h, s19 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v11.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v10.l, v26.l, s20 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s19, 0, v36 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v26.l, v10.h, s21 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0x8000, v10.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.l, v11.l, v26.h, s19 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s20, v10, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v36.h, v10.h, s20 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s20, 0x8000, v36.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v10.l, v10.h, s21 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s21, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v8 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v25.h, v9.h, v36.h, s20 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v10.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v9.l, v25.l, s21 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s20, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v25.l, v9.h, s22 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0x8000, v9.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v1.h, v4.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v8.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v16.l, v2.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v5.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v2.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v7.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.l, v10.l, v25.h, s20 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s21, v9, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v36.h, v9.h, s21 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s21, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v9.l, v9.h, s22 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s22, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v7 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v24.h, v8.h, v36.h, s21 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v9.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v8.l, v24.l, s22 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s21, 0, v36 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v24.l, v8.h, s23 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0x8000, v8.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.l, v9.l, v24.h, s21 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s22, v8, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v36.h, v8.h, s22 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s22, 0x8000, v36.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v8.l, v8.h, s23 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s23, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v23.h, v7.h, v36.h, s22 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v8.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v7.l, v23.l, s23 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s22, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v23.l, v7.h, s24 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0x8000, v7.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v7.l, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v8.l, v23.h, s22 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s23, v7, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v36.h, v7.h, s23 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s23, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v7.l, v7.h, s24 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s24, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v5 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v22.h, v6.h, v36.h, s23 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v7.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v6.l, v22.l, s24 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s23, 0, v36 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v22.l, v6.h, s25 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0x8000, v6.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v7.l, v22.h, s23 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s24, v6, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v36.h, v6.h, s24 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s24, 0x8000, v36.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v6.l, v6.h, s25 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s25, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v12 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v16 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v11 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s3, 0, v10 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v29.l, v5.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v21.h, v5.h, v36.h, s24 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v6.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v5.l, v21.l, s25 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v36.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s24, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v21.l, v5.h, s26 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0x8000, v5.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v49.l, v2.l, v1.l, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v48.l, v8.l, v1.h, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v39.l, v6.l, v0.h, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v38.l, v9.l, v2.h, s3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v0, v29 :: v_dual_mov_b32 v1, v49 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v2, v48 :: v_dual_mov_b32 v3, v39 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v4, v38 :: v_dual_mov_b32 v5, v36 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v6, v35 :: v_dual_mov_b32 v7, v34 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v8, v33 :: v_dual_mov_b32 v9, v32 -; GFX12-TRUE16-NEXT: v_dual_mov_b32 v10, v31 :: v_dual_mov_b32 v11, v30 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v12, v37 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v6.l, v21.h, s24 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s25, v5, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v36.h, v5.h, s25 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s25, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v5.l, v5.h, s26 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s26, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v20.h, v4.h, v36.h, s25 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v5.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v20.l, s26 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s25, 0, v36 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v20.l, v4.h, s27 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s27, 0x8000, v4.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v5.l, v20.h, s25 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s26, v4, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v36.h, v4.h, s26 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s26, 0x8000, v36.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v4.l, v4.h, s27 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v81, v81 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v81, 16, v18 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v18.h, v3.h, v36.h, s26 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v4.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v83.h, v3.l, v19.l, s27 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s29, v81, v81 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s26, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v19.l, v83.h, s28 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s28, v84, v84 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s40, v3, v3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s41, 0x8000, v83.h +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v19.l, v36.l +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s27, v83, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v82.h, v2.l, v18.l, s28 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v19.h, v1.l, v17.l, s40 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v32.h, v37.h, s44 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v36.h, v83.h, s27 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s40, 0x8000, v82.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s27, v85, v85 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.l, v18.h, s26 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v14.h, v32.l, v1.l, s0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v3.l, v83.h, s41 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s41, v87, v87 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.h, v36.h, s28 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v3.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v31.h, v35.h, s42 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s42, 0x8000, v39.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s28, 0x8000, v19.h +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s45, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v18.l, v82.h, s29 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s29, 0x8000, v38.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.h, v31.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v34.h, v39.h, s42 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v48.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s44, v82, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v33.h, v38.h, s29 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v12.h, v34.l, v1.h, s2 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v37.l, v48.h, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v36.h, v82.h, s44 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v13.h, v33.l, v0.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v49.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v50.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v11.h, v35.l, v1.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v82.h, s40 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v51.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v54.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v49.l, v50.h, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v0.l, s45 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v36.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v1.l +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v39.l, v49.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v52.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v51.l, v51.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v53.h +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v17.l, v19.h, s27 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v10.h, v38.l, v0.h, s4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v69.l, v52.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v55.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v9.h, v48.l, v1.h, s5 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v19, v36 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v36.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v69.h, v53.h, s2 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v70.h, v55.h, s4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v66.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.h, v36.h, v19.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v65.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v50.l, v2.h, s6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v70.l, v54.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v52.l, v0.h, s7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.l, v16.h, v19.h, s28 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v53.l, v1.h, s8 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v71.h, v65.h, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v54.l, v2.h, s9 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v67.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.l, v17.l, v36.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v16.h +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v64.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v65.l, v1.h, s12 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v15.l, v67.h, s3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v55.l, v3.h, s10 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s4, 0, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v36.h, v16.l, v96.h, s41 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v71.l, v64.h, s0 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.l, v80.l, v66.h, s2 +; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v2.l, vcc_lo +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v16.h, v17.l, s4 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v96, v36 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v64.l, v0.h, s11 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v66.l, v16.l, s13 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v67.l, v15.l, s14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v15.l, v68.l, v30.h, s15 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v16.l, v36.h, v96.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v36.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.h, v16.l, v96.h, s43 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v17.h, v17.h, v36.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v36.h, v16.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v36 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v16.l, v17.h, s0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v32bf16: @@ -14669,34 +14424,29 @@ define bfloat @v_minimumnum_bf16_no_ieee(bfloat %x, bfloat %y) #0 { ; GFX11-TRUE16-LABEL: v_minimumnum_bf16_no_ieee: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v2, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v1, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_bf16_no_ieee: @@ -14732,40 +14482,34 @@ define bfloat @v_minimumnum_bf16_no_ieee(bfloat %x, bfloat %y) #0 { ; GFX12-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v1.l +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.h, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v2, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v0.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v1.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v1.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v1, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v1.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v1.h, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, vcc_lo +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v2.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_bf16_no_ieee: @@ -15009,58 +14753,46 @@ define <2 x bfloat> @v_minimumnum_v2bf16_no_ieee(<2 x bfloat> %x, <2 x bfloat> % ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v5, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v1.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v4.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v4, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.h, v4.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v4.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v1.l, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.l, v3.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v3, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v3.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v3.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v2.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v0.h, vcc_lo ; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v2.l, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v4, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v1.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v2bf16_no_ieee: @@ -15117,62 +14849,56 @@ define <2 x bfloat> @v_minimumnum_v2bf16_no_ieee(<2 x bfloat> %x, <2 x bfloat> % ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v4, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v5, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v1.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v0.h, v1.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v2.l, s1 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v0.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.l -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v4, v6 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v4.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v4, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v3.l, v2.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.h, v4.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v1.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v7 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v3.l, s1 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.h, v0.l, v1.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v1.l, v3.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v3, v4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v4.h, v3.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v0.l, v3.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v2.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v4.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v2.h, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v2.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v5.l, v0.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v1.l, s0 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v2bf16_no_ieee: @@ -15521,66 +15247,67 @@ define <3 x bfloat> @v_minimumnum_v3bf16_no_ieee(<3 x bfloat> %x, <3 x bfloat> % ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v0 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v6, v6 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.h, v4.l, s1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v8 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v7, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v9, v11 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v2.l, v0.l, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v4.l, s0 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s0 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v7, v7 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v7, v7 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v4 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v5.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v5.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v5.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v3, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v4.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.h, v4.h, s2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v3bf16_no_ieee: @@ -15652,77 +15379,80 @@ define <3 x bfloat> @v_minimumnum_v3bf16_no_ieee(<3 x bfloat> %x, <3 x bfloat> % ; GFX12-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v6, v6 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v5, v5 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v0.h, v2.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v9, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v2.h, v4.l, s1 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v7, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v8 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v4.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v7, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v1.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s3, v9, v11 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v2.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v7, v7 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v5.l, v4.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v3.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v2.l, v0.l, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v4.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v7.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v8.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v4.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v4 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v1.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v5.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v5.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v5, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v5.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v5.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v3, v3 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v0.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s0 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v5 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v9 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v6.l, v0.h, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v0, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s1 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v4.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.h, v4.h, s2 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v8.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v0.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v7.l, v1.l, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s1 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v3bf16_no_ieee: @@ -16183,85 +15913,83 @@ define <4 x bfloat> @v_minimumnum_v4bf16_no_ieee(<4 x bfloat> %x, <4 x bfloat> % ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v0 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v0.h, v2.h, s1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v4.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v10, v10 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v1.h, v3.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s4 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, s3 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v10, v8 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v0.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.l, v5.l, v4.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v9, v10 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v12 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, s1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v7.l, v6.l, s2 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v8.l -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v11, v12 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v7.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v9, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v14 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v1.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v7.l, s2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.l -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s1 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v10 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v0.h, s0 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.h, s2 -; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v8, v8 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v7.h, v2.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v7 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v7.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.h, v6.h, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v5.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v8.h, v1.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v8.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v8, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v8.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v8.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v3, v3 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v7.h, s4 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX11-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v0, v6 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v6.h +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s3 +; GFX11-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v5.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s2 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v5.h, s3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v0.h, s1 +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: v_minimumnum_v4bf16_no_ieee: @@ -16356,99 +16084,98 @@ define <4 x bfloat> @v_minimumnum_v4bf16_no_ieee(<4 x bfloat> %x, <4 x bfloat> % ; GFX12-TRUE16-NEXT: s_wait_kmcnt 0x0 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 ; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v0 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v2 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v5, v5 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s1, v6, v6 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v0 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v8, v8 -; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v1.h, v3.h, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v0.h, v2.h, s1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.l ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v3 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v2 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.h, v4.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s4, v10, v10 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v11, v11 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v7 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.l, v2.h, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s2 -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s4 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v4.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v3.l, v1.l, s3 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v2.l, v0.l, s0 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v10, v8 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v11.l, v1.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v12.l, v0.l -; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.l, v5.l, v4.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v15.l, v2.l -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v6.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11 -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v9, v10 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v12 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v8.l, v4.l, s1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v15 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v1.h, v3.h, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v9, v9 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v7.l, v6.l, s2 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v5.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v14.l, v8.l -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s1, v11, v12 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v7.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.h, v3.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s3, v9, v9 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v6.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v9, v13 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v14 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.l, v5.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.h, v5.l, s0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v3.l, v1.l, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v4.h, v4.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s0, v8, v8 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v1 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v6 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.l, v2.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v1.l -; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v1.h, v7.l, s2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v0.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v0.h, v2.h, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s1, 0x8000, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v7.h, v2.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e32 vcc_lo, v6, v7 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s4, 0x8000, v7.h ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e32 vcc_lo, 0x8000, v2.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v7 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v5.l, v7.h, v6.h, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v3.l, s1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v5.l, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v5.l ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, vcc_lo -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v9 -; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s1, 0, v10 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v8.l, v0.h, s0 -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.l, v2.h, s2 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v8.h, v1.l, v3.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v3.l, v8.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v8.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0 +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s0, v8, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v6.h, v8.h, s0 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s0, 0x8000, v6.h +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v1.l, v8.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e64 s2, v3, v3 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v3.l, v1.h, v7.h, s4 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.h, v0.h, v6.h, s0 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v1.l +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v2.l, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v0.l, v6.l +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s0, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v6.h, v2.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v0.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v1.l, v2.h, s0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_cmp_lt_f32_e64 s2, v0, v6 +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.h, v0.h, s2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s2, 0x8000, v6.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v0.l, v0.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_u16_e64 s3, 0x8000, v5.h +; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff +; GFX12-TRUE16-NEXT: v_cndmask_b16 v2.l, v0.h, v6.h, s2 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, v0.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v4.h, v5.h, s3 +; GFX12-TRUE16-NEXT: v_cmp_eq_f32_e64 s2, 0, v6 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.h, v4.l, v0.h, s1 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v6.l, v0.l, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.h, v5.l, v3.l, vcc_lo ; GFX12-TRUE16-NEXT: s_wait_alu 0xf1ff -; GFX12-TRUE16-NEXT: v_cndmask_b16 v1.l, v5.l, v1.l, s1 +; GFX12-TRUE16-NEXT: v_cndmask_b16 v0.l, v0.l, v2.l, s2 ; GFX12-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX12-FAKE16-LABEL: v_minimumnum_v4bf16_no_ieee: diff --git a/llvm/test/CodeGen/AMDGPU/sub.v2i16.ll b/llvm/test/CodeGen/AMDGPU/sub.v2i16.ll index 42bd2ff..9f539bd 100644 --- a/llvm/test/CodeGen/AMDGPU/sub.v2i16.ll +++ b/llvm/test/CodeGen/AMDGPU/sub.v2i16.ll @@ -813,7 +813,8 @@ define amdgpu_kernel void @v_test_sub_v2i16_zext_to_v2i64(ptr addrspace(1) %out, ; GFX11-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 ; GFX11-TRUE16-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0x3ff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 2, v0 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: global_load_b32 v1, v0, s[2:3] glc dlc @@ -824,11 +825,12 @@ define amdgpu_kernel void @v_test_sub_v2i16_zext_to_v2i64(ptr addrspace(1) %out, ; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1 ; GFX11-TRUE16-NEXT: v_pk_sub_i16 v0, v1, v0 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX11-TRUE16-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_and_b32 v0, 0xffff, v0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshl_or_b32 v2, 0, 16, v2 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, 16, v0 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l +; GFX11-TRUE16-NEXT: v_lshl_or_b32 v2, v2, 16, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v1 ; GFX11-TRUE16-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_endpgm ; diff --git a/llvm/test/CodeGen/AMDGPU/vector-reduce-umax.ll b/llvm/test/CodeGen/AMDGPU/vector-reduce-umax.ll index 184c807..ddae1b2 100644 --- a/llvm/test/CodeGen/AMDGPU/vector-reduce-umax.ll +++ b/llvm/test/CodeGen/AMDGPU/vector-reduce-umax.ll @@ -1718,7 +1718,7 @@ define i16 @test_vector_reduce_umax_v3i16(<3 x i16> %v) { ; GFX11-SDAG-TRUE16-LABEL: test_vector_reduce_umax_v3i16: ; GFX11-SDAG-TRUE16: ; %bb.0: ; %entry ; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 +; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 ; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-SDAG-TRUE16-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX11-SDAG-TRUE16-NEXT: v_max_u16 v0.l, v0.l, v0.h @@ -1751,7 +1751,7 @@ define i16 @test_vector_reduce_umax_v3i16(<3 x i16> %v) { ; GFX12-SDAG-TRUE16-NEXT: s_wait_samplecnt 0x0 ; GFX12-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0 ; GFX12-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0 -; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 +; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 ; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-SDAG-TRUE16-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX12-SDAG-TRUE16-NEXT: v_max_u16 v0.l, v0.l, v0.h diff --git a/llvm/test/CodeGen/NVPTX/i8x4-instructions.ll b/llvm/test/CodeGen/NVPTX/i8x4-instructions.ll index cbc9f70..aba20e6 100644 --- a/llvm/test/CodeGen/NVPTX/i8x4-instructions.ll +++ b/llvm/test/CodeGen/NVPTX/i8x4-instructions.ll @@ -2311,4 +2311,51 @@ entry: ret void } +define <4 x float> @test_uitofp_v4i8(<4 x i8> %a) { +; CHECK-LABEL: test_uitofp_v4i8( +; CHECK: { +; CHECK-NEXT: .reg .b32 %r<10>; +; CHECK-EMPTY: +; CHECK-NEXT: // %bb.0: +; CHECK-NEXT: ld.param.b32 %r1, [test_uitofp_v4i8_param_0]; +; CHECK-NEXT: prmt.b32 %r2, %r1, 0, 0x7773U; +; CHECK-NEXT: cvt.rn.f32.u32 %r3, %r2; +; CHECK-NEXT: prmt.b32 %r4, %r1, 0, 0x7772U; +; CHECK-NEXT: cvt.rn.f32.u32 %r5, %r4; +; CHECK-NEXT: prmt.b32 %r6, %r1, 0, 0x7771U; +; CHECK-NEXT: cvt.rn.f32.u32 %r7, %r6; +; CHECK-NEXT: prmt.b32 %r8, %r1, 0, 0x7770U; +; CHECK-NEXT: cvt.rn.f32.u32 %r9, %r8; +; CHECK-NEXT: st.param.v4.b32 [func_retval0], {%r9, %r7, %r5, %r3}; +; CHECK-NEXT: ret; + %r = uitofp <4 x i8> %a to <4 x float> + ret <4 x float> %r +} + +define <4 x float> @test_sitofp_v4i8(<4 x i8> %a) { +; CHECK-LABEL: test_sitofp_v4i8( +; CHECK: { +; CHECK-NEXT: .reg .b16 %rs<5>; +; CHECK-NEXT: .reg .b32 %r<10>; +; CHECK-EMPTY: +; CHECK-NEXT: // %bb.0: +; CHECK-NEXT: ld.param.b32 %r1, [test_sitofp_v4i8_param_0]; +; CHECK-NEXT: prmt.b32 %r2, %r1, 0, 0xbbb3U; +; CHECK-NEXT: cvt.u16.u32 %rs1, %r2; +; CHECK-NEXT: cvt.rn.f32.s16 %r3, %rs1; +; CHECK-NEXT: prmt.b32 %r4, %r1, 0, 0xaaa2U; +; CHECK-NEXT: cvt.u16.u32 %rs2, %r4; +; CHECK-NEXT: cvt.rn.f32.s16 %r5, %rs2; +; CHECK-NEXT: prmt.b32 %r6, %r1, 0, 0x9991U; +; CHECK-NEXT: cvt.u16.u32 %rs3, %r6; +; CHECK-NEXT: cvt.rn.f32.s16 %r7, %rs3; +; CHECK-NEXT: prmt.b32 %r8, %r1, 0, 0x8880U; +; CHECK-NEXT: cvt.u16.u32 %rs4, %r8; +; CHECK-NEXT: cvt.rn.f32.s16 %r9, %rs4; +; CHECK-NEXT: st.param.v4.b32 [func_retval0], {%r9, %r7, %r5, %r3}; +; CHECK-NEXT: ret; + %r = sitofp <4 x i8> %a to <4 x float> + ret <4 x float> %r +} + attributes #0 = { nounwind } diff --git a/llvm/test/CodeGen/NVPTX/trunc-tofp.ll b/llvm/test/CodeGen/NVPTX/trunc-tofp.ll new file mode 100644 index 0000000..404c423 --- /dev/null +++ b/llvm/test/CodeGen/NVPTX/trunc-tofp.ll @@ -0,0 +1,81 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 +; RUN: llc < %s -mcpu=sm_80 | FileCheck %s +; RUN: %if ptxas %{ llc < %s -mcpu=sm_80 | %ptxas-verify %} + +target triple = "nvptx64-nvidia-cuda" + +define float @uitofp_trunc_nuw(i32 %x, i32 %y) { +; CHECK-LABEL: uitofp_trunc_nuw( +; CHECK: { +; CHECK-NEXT: .reg .b32 %r<5>; +; CHECK-EMPTY: +; CHECK-NEXT: // %bb.0: +; CHECK-NEXT: ld.param.b32 %r1, [uitofp_trunc_nuw_param_0]; +; CHECK-NEXT: ld.param.b32 %r2, [uitofp_trunc_nuw_param_1]; +; CHECK-NEXT: add.s32 %r3, %r1, %r2; +; CHECK-NEXT: cvt.rn.f32.u32 %r4, %r3; +; CHECK-NEXT: st.param.b32 [func_retval0], %r4; +; CHECK-NEXT: ret; + %v = add i32 %x, %y + %t = trunc nuw i32 %v to i16 + %f = uitofp i16 %t to float + ret float %f +} + +define float @sitofp_trunc_nsw(i32 %x, i32 %y) { +; CHECK-LABEL: sitofp_trunc_nsw( +; CHECK: { +; CHECK-NEXT: .reg .b32 %r<5>; +; CHECK-EMPTY: +; CHECK-NEXT: // %bb.0: +; CHECK-NEXT: ld.param.b32 %r1, [sitofp_trunc_nsw_param_0]; +; CHECK-NEXT: ld.param.b32 %r2, [sitofp_trunc_nsw_param_1]; +; CHECK-NEXT: add.s32 %r3, %r1, %r2; +; CHECK-NEXT: cvt.rn.f32.s32 %r4, %r3; +; CHECK-NEXT: st.param.b32 [func_retval0], %r4; +; CHECK-NEXT: ret; + %v = add i32 %x, %y + %t = trunc nsw i32 %v to i16 + %f = sitofp i16 %t to float + ret float %f +} + +define float @uitofp_trunc_nsw(i32 %x, i32 %y) { +; CHECK-LABEL: uitofp_trunc_nsw( +; CHECK: { +; CHECK-NEXT: .reg .b16 %rs<2>; +; CHECK-NEXT: .reg .b32 %r<5>; +; CHECK-EMPTY: +; CHECK-NEXT: // %bb.0: +; CHECK-NEXT: ld.param.b32 %r1, [uitofp_trunc_nsw_param_0]; +; CHECK-NEXT: ld.param.b32 %r2, [uitofp_trunc_nsw_param_1]; +; CHECK-NEXT: add.s32 %r3, %r1, %r2; +; CHECK-NEXT: cvt.u16.u32 %rs1, %r3; +; CHECK-NEXT: cvt.rn.f32.u16 %r4, %rs1; +; CHECK-NEXT: st.param.b32 [func_retval0], %r4; +; CHECK-NEXT: ret; + %v = add i32 %x, %y + %t = trunc nsw i32 %v to i16 + %f = uitofp i16 %t to float + ret float %f +} + +define float @sitofp_trunc_nuw(i32 %x, i32 %y) { +; CHECK-LABEL: sitofp_trunc_nuw( +; CHECK: { +; CHECK-NEXT: .reg .b16 %rs<2>; +; CHECK-NEXT: .reg .b32 %r<5>; +; CHECK-EMPTY: +; CHECK-NEXT: // %bb.0: +; CHECK-NEXT: ld.param.b32 %r1, [sitofp_trunc_nuw_param_0]; +; CHECK-NEXT: ld.param.b32 %r2, [sitofp_trunc_nuw_param_1]; +; CHECK-NEXT: add.s32 %r3, %r1, %r2; +; CHECK-NEXT: cvt.u16.u32 %rs1, %r3; +; CHECK-NEXT: cvt.rn.f32.s16 %r4, %rs1; +; CHECK-NEXT: st.param.b32 [func_retval0], %r4; +; CHECK-NEXT: ret; + %v = add i32 %x, %y + %t = trunc nuw i32 %v to i16 + %f = sitofp i16 %t to float + ret float %f +} diff --git a/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-load.ll b/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-load.ll index 9af92aa..578b67e 100644 --- a/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-load.ll +++ b/llvm/test/CodeGen/RISCV/rvv/vector-deinterleave-load.ll @@ -538,3 +538,164 @@ define { <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x %res7 = insertvalue { <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8> } %res6, <vscale x 8 x i8> %t7, 7 ret { <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8> } %res7 } + +define {<vscale x 16 x i8>, <vscale x 16 x i8>} @masked_load_factor2(ptr %p) { +; CHECK-LABEL: masked_load_factor2: +; CHECK: # %bb.0: +; CHECK-NEXT: vl4r.v v12, (a0) +; CHECK-NEXT: vsetvli a0, zero, e8, m2, ta, ma +; CHECK-NEXT: vnsrl.wi v8, v12, 0 +; CHECK-NEXT: vnsrl.wi v10, v12, 8 +; CHECK-NEXT: ret + %vec = call <vscale x 32 x i8> @llvm.masked.load(ptr %p, i32 4, <vscale x 32 x i1> splat (i1 true), <vscale x 32 x i8> poison) + %deinterleaved.results = call {<vscale x 16 x i8>, <vscale x 16 x i8>} @llvm.vector.deinterleave2.nxv32i8(<vscale x 32 x i8> %vec) + ret {<vscale x 16 x i8>, <vscale x 16 x i8>} %deinterleaved.results +} + +define {<vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>} @masked_loat_factor4(ptr %p) { +; CHECK-LABEL: masked_loat_factor4: +; CHECK: # %bb.0: +; CHECK-NEXT: addi sp, sp, -16 +; CHECK-NEXT: .cfi_def_cfa_offset 16 +; CHECK-NEXT: csrr a1, vlenb +; CHECK-NEXT: slli a1, a1, 2 +; CHECK-NEXT: sub sp, sp, a1 +; CHECK-NEXT: .cfi_escape 0x0f, 0x0d, 0x72, 0x00, 0x11, 0x10, 0x22, 0x11, 0x04, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 16 + 4 * vlenb +; CHECK-NEXT: vl4r.v v8, (a0) +; CHECK-NEXT: addi a0, sp, 16 +; CHECK-NEXT: vs4r.v v8, (a0) +; CHECK-NEXT: vsetvli a1, zero, e8, m1, ta, ma +; CHECK-NEXT: vlseg4e8.v v8, (a0) +; CHECK-NEXT: csrr a0, vlenb +; CHECK-NEXT: slli a0, a0, 2 +; CHECK-NEXT: add sp, sp, a0 +; CHECK-NEXT: .cfi_def_cfa sp, 16 +; CHECK-NEXT: addi sp, sp, 16 +; CHECK-NEXT: .cfi_def_cfa_offset 0 +; CHECK-NEXT: ret + %vec = call <vscale x 32 x i8> @llvm.masked.load(ptr %p, i32 4, <vscale x 32 x i1> splat (i1 true), <vscale x 32 x i8> poison) + %deinterleaved.results = call {<vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>} @llvm.vector.deinterleave4.nxv32i8(<vscale x 32 x i8> %vec) + ret {<vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>} %deinterleaved.results +} + +define {<vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>} @masked_loat_factor4_mask(ptr %p, <vscale x 8 x i1> %mask) { +; CHECK-LABEL: masked_loat_factor4_mask: +; CHECK: # %bb.0: +; CHECK-NEXT: addi sp, sp, -16 +; CHECK-NEXT: .cfi_def_cfa_offset 16 +; CHECK-NEXT: csrr a1, vlenb +; CHECK-NEXT: slli a1, a1, 3 +; CHECK-NEXT: sub sp, sp, a1 +; CHECK-NEXT: .cfi_escape 0x0f, 0x0d, 0x72, 0x00, 0x11, 0x10, 0x22, 0x11, 0x08, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 16 + 8 * vlenb +; CHECK-NEXT: vsetvli a1, zero, e8, m1, ta, ma +; CHECK-NEXT: vmv.v.i v8, 0 +; CHECK-NEXT: addi a1, sp, 16 +; CHECK-NEXT: csrr a2, vlenb +; CHECK-NEXT: vmerge.vim v8, v8, 1, v0 +; CHECK-NEXT: add a3, a1, a2 +; CHECK-NEXT: vmv.v.v v9, v8 +; CHECK-NEXT: srli a4, a2, 2 +; CHECK-NEXT: vmv.v.v v10, v8 +; CHECK-NEXT: srli a5, a2, 3 +; CHECK-NEXT: vmv.v.v v11, v8 +; CHECK-NEXT: vsseg4e8.v v8, (a1) +; CHECK-NEXT: vl1r.v v8, (a1) +; CHECK-NEXT: add a1, a4, a5 +; CHECK-NEXT: vl1r.v v9, (a3) +; CHECK-NEXT: add a3, a3, a2 +; CHECK-NEXT: add a2, a3, a2 +; CHECK-NEXT: vl1r.v v10, (a3) +; CHECK-NEXT: vl1r.v v11, (a2) +; CHECK-NEXT: vmsne.vi v9, v9, 0 +; CHECK-NEXT: vmsne.vi v0, v8, 0 +; CHECK-NEXT: vmsne.vi v8, v10, 0 +; CHECK-NEXT: vmsne.vi v10, v11, 0 +; CHECK-NEXT: vsetvli zero, a4, e8, mf2, tu, ma +; CHECK-NEXT: vslideup.vx v0, v9, a5 +; CHECK-NEXT: vsetvli zero, a1, e8, mf2, tu, ma +; CHECK-NEXT: vslideup.vx v0, v8, a4 +; CHECK-NEXT: vsetvli a2, zero, e8, mf2, ta, ma +; CHECK-NEXT: vslideup.vx v0, v10, a1 +; CHECK-NEXT: vsetvli a1, zero, e8, m4, ta, ma +; CHECK-NEXT: vle8.v v8, (a0), v0.t +; CHECK-NEXT: csrr a0, vlenb +; CHECK-NEXT: slli a0, a0, 2 +; CHECK-NEXT: add a0, sp, a0 +; CHECK-NEXT: addi a0, a0, 16 +; CHECK-NEXT: vs4r.v v8, (a0) +; CHECK-NEXT: vsetvli a1, zero, e8, m1, ta, ma +; CHECK-NEXT: vlseg4e8.v v8, (a0) +; CHECK-NEXT: csrr a0, vlenb +; CHECK-NEXT: slli a0, a0, 3 +; CHECK-NEXT: add sp, sp, a0 +; CHECK-NEXT: .cfi_def_cfa sp, 16 +; CHECK-NEXT: addi sp, sp, 16 +; CHECK-NEXT: .cfi_def_cfa_offset 0 +; CHECK-NEXT: ret + %interleaved.mask = tail call <vscale x 32 x i1> @llvm.vector.interleave4.nxv32i1(<vscale x 8 x i1> %mask, <vscale x 8 x i1> %mask, <vscale x 8 x i1> %mask, <vscale x 8 x i1> %mask) + %vec = call <vscale x 32 x i8> @llvm.masked.load(ptr %p, i32 4, <vscale x 32 x i1> %interleaved.mask, <vscale x 32 x i8> poison) + %deinterleaved.results = call {<vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>} @llvm.vector.deinterleave4.nxv32i8(<vscale x 32 x i8> %vec) + ret {<vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>} %deinterleaved.results +} + +; Negative test - some of the deinterleaved elements might come from the +; passthru not the load +define {<vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>} @masked_loat_factor4_passthru(ptr %p, <vscale x 8 x i1> %mask, <vscale x 32 x i8> %passthru) { +; CHECK-LABEL: masked_loat_factor4_passthru: +; CHECK: # %bb.0: +; CHECK-NEXT: addi sp, sp, -16 +; CHECK-NEXT: .cfi_def_cfa_offset 16 +; CHECK-NEXT: csrr a1, vlenb +; CHECK-NEXT: slli a1, a1, 3 +; CHECK-NEXT: sub sp, sp, a1 +; CHECK-NEXT: .cfi_escape 0x0f, 0x0d, 0x72, 0x00, 0x11, 0x10, 0x22, 0x11, 0x08, 0x92, 0xa2, 0x38, 0x00, 0x1e, 0x22 # sp + 16 + 8 * vlenb +; CHECK-NEXT: vsetvli a1, zero, e8, m1, ta, ma +; CHECK-NEXT: vmv.v.i v12, 0 +; CHECK-NEXT: addi a1, sp, 16 +; CHECK-NEXT: csrr a2, vlenb +; CHECK-NEXT: vmerge.vim v12, v12, 1, v0 +; CHECK-NEXT: add a3, a1, a2 +; CHECK-NEXT: vmv.v.v v13, v12 +; CHECK-NEXT: srli a4, a2, 2 +; CHECK-NEXT: vmv.v.v v14, v12 +; CHECK-NEXT: srli a5, a2, 3 +; CHECK-NEXT: vmv.v.v v15, v12 +; CHECK-NEXT: vsseg4e8.v v12, (a1) +; CHECK-NEXT: vl1r.v v12, (a1) +; CHECK-NEXT: add a1, a4, a5 +; CHECK-NEXT: vl1r.v v13, (a3) +; CHECK-NEXT: add a3, a3, a2 +; CHECK-NEXT: add a2, a3, a2 +; CHECK-NEXT: vl1r.v v14, (a3) +; CHECK-NEXT: vl1r.v v15, (a2) +; CHECK-NEXT: vmsne.vi v13, v13, 0 +; CHECK-NEXT: vmsne.vi v0, v12, 0 +; CHECK-NEXT: vmsne.vi v12, v14, 0 +; CHECK-NEXT: vmsne.vi v14, v15, 0 +; CHECK-NEXT: vsetvli zero, a4, e8, mf2, tu, ma +; CHECK-NEXT: vslideup.vx v0, v13, a5 +; CHECK-NEXT: vsetvli zero, a1, e8, mf2, tu, ma +; CHECK-NEXT: vslideup.vx v0, v12, a4 +; CHECK-NEXT: vsetvli a2, zero, e8, mf2, ta, ma +; CHECK-NEXT: vslideup.vx v0, v14, a1 +; CHECK-NEXT: vsetvli a1, zero, e8, m4, ta, mu +; CHECK-NEXT: vle8.v v8, (a0), v0.t +; CHECK-NEXT: csrr a0, vlenb +; CHECK-NEXT: slli a0, a0, 2 +; CHECK-NEXT: add a0, sp, a0 +; CHECK-NEXT: addi a0, a0, 16 +; CHECK-NEXT: vs4r.v v8, (a0) +; CHECK-NEXT: vsetvli a1, zero, e8, m1, ta, ma +; CHECK-NEXT: vlseg4e8.v v8, (a0) +; CHECK-NEXT: csrr a0, vlenb +; CHECK-NEXT: slli a0, a0, 3 +; CHECK-NEXT: add sp, sp, a0 +; CHECK-NEXT: .cfi_def_cfa sp, 16 +; CHECK-NEXT: addi sp, sp, 16 +; CHECK-NEXT: .cfi_def_cfa_offset 0 +; CHECK-NEXT: ret + %interleaved.mask = tail call <vscale x 32 x i1> @llvm.vector.interleave4.nxv32i1(<vscale x 8 x i1> %mask, <vscale x 8 x i1> %mask, <vscale x 8 x i1> %mask, <vscale x 8 x i1> %mask) + %vec = call <vscale x 32 x i8> @llvm.masked.load(ptr %p, i32 4, <vscale x 32 x i1> %interleaved.mask, <vscale x 32 x i8> %passthru) + %deinterleaved.results = call {<vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>} @llvm.vector.deinterleave4.nxv32i8(<vscale x 32 x i8> %vec) + ret {<vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>, <vscale x 8 x i8>} %deinterleaved.results +} diff --git a/llvm/test/CodeGen/RISCV/short-forward-branch-opt.ll b/llvm/test/CodeGen/RISCV/short-forward-branch-opt.ll index 061435c..59a702a 100644 --- a/llvm/test/CodeGen/RISCV/short-forward-branch-opt.ll +++ b/llvm/test/CodeGen/RISCV/short-forward-branch-opt.ll @@ -798,12 +798,12 @@ define void @sextw_removal_ccor(i1 %c, i32 signext %arg, i32 signext %arg1, i32 ; RV64SFBSIFIVEU74-LABEL: sextw_removal_ccor: ; RV64SFBSIFIVEU74: # %bb.0: # %bb ; RV64SFBSIFIVEU74-NEXT: addi sp, sp, -32 -; RV64SFBSIFIVEU74-NEXT: sd ra, 24(sp) # 8-byte Folded Spill ; RV64SFBSIFIVEU74-NEXT: sd s0, 16(sp) # 8-byte Folded Spill -; RV64SFBSIFIVEU74-NEXT: sd s1, 8(sp) # 8-byte Folded Spill ; RV64SFBSIFIVEU74-NEXT: mv s0, a3 +; RV64SFBSIFIVEU74-NEXT: sd s1, 8(sp) # 8-byte Folded Spill ; RV64SFBSIFIVEU74-NEXT: andi a0, a0, 1 ; RV64SFBSIFIVEU74-NEXT: mv s1, a2 +; RV64SFBSIFIVEU74-NEXT: sd ra, 24(sp) # 8-byte Folded Spill ; RV64SFBSIFIVEU74-NEXT: beqz a0, .LBB15_4 ; RV64SFBSIFIVEU74-NEXT: # %bb.3: # %bb ; RV64SFBSIFIVEU74-NEXT: or s0, a3, a1 @@ -824,11 +824,11 @@ define void @sextw_removal_ccor(i1 %c, i32 signext %arg, i32 signext %arg1, i32 ; RV64SFBANDESAX45-LABEL: sextw_removal_ccor: ; RV64SFBANDESAX45: # %bb.0: # %bb ; RV64SFBANDESAX45-NEXT: addi sp, sp, -32 -; RV64SFBANDESAX45-NEXT: sd ra, 24(sp) # 8-byte Folded Spill ; RV64SFBANDESAX45-NEXT: sd s0, 16(sp) # 8-byte Folded Spill -; RV64SFBANDESAX45-NEXT: sd s1, 8(sp) # 8-byte Folded Spill ; RV64SFBANDESAX45-NEXT: mv s0, a3 +; RV64SFBANDESAX45-NEXT: sd s1, 8(sp) # 8-byte Folded Spill ; RV64SFBANDESAX45-NEXT: mv s1, a2 +; RV64SFBANDESAX45-NEXT: sd ra, 24(sp) # 8-byte Folded Spill ; RV64SFBANDESAX45-NEXT: nds.bbc a0, 0, .LBB15_2 ; RV64SFBANDESAX45-NEXT: # %bb.1: ; RV64SFBANDESAX45-NEXT: or s0, s0, a1 @@ -848,12 +848,12 @@ define void @sextw_removal_ccor(i1 %c, i32 signext %arg, i32 signext %arg1, i32 ; ZICOND-LABEL: sextw_removal_ccor: ; ZICOND: # %bb.0: # %bb ; ZICOND-NEXT: addi sp, sp, -32 -; ZICOND-NEXT: sd ra, 24(sp) # 8-byte Folded Spill ; ZICOND-NEXT: sd s0, 16(sp) # 8-byte Folded Spill -; ZICOND-NEXT: sd s1, 8(sp) # 8-byte Folded Spill ; ZICOND-NEXT: mv s0, a3 +; ZICOND-NEXT: sd s1, 8(sp) # 8-byte Folded Spill ; ZICOND-NEXT: andi a0, a0, 1 ; ZICOND-NEXT: mv s1, a2 +; ZICOND-NEXT: sd ra, 24(sp) # 8-byte Folded Spill ; ZICOND-NEXT: beqz a0, .LBB15_4 ; ZICOND-NEXT: # %bb.3: # %bb ; ZICOND-NEXT: or s0, a3, a1 @@ -874,12 +874,12 @@ define void @sextw_removal_ccor(i1 %c, i32 signext %arg, i32 signext %arg1, i32 ; RV32SFB-LABEL: sextw_removal_ccor: ; RV32SFB: # %bb.0: # %bb ; RV32SFB-NEXT: addi sp, sp, -16 -; RV32SFB-NEXT: sw ra, 12(sp) # 4-byte Folded Spill ; RV32SFB-NEXT: sw s0, 8(sp) # 4-byte Folded Spill -; RV32SFB-NEXT: sw s1, 4(sp) # 4-byte Folded Spill ; RV32SFB-NEXT: mv s0, a3 +; RV32SFB-NEXT: sw s1, 4(sp) # 4-byte Folded Spill ; RV32SFB-NEXT: andi a0, a0, 1 ; RV32SFB-NEXT: mv s1, a2 +; RV32SFB-NEXT: sw ra, 12(sp) # 4-byte Folded Spill ; RV32SFB-NEXT: beqz a0, .LBB15_4 ; RV32SFB-NEXT: # %bb.3: # %bb ; RV32SFB-NEXT: or s0, a3, a1 @@ -941,11 +941,11 @@ define void @sextw_removal_ccaddw(i1 %c, i32 signext %arg, i32 signext %arg1, i3 ; RV64SFBSIFIVEU74-LABEL: sextw_removal_ccaddw: ; RV64SFBSIFIVEU74: # %bb.0: # %bb ; RV64SFBSIFIVEU74-NEXT: addi sp, sp, -32 -; RV64SFBSIFIVEU74-NEXT: sd ra, 24(sp) # 8-byte Folded Spill -; RV64SFBSIFIVEU74-NEXT: sd s0, 16(sp) # 8-byte Folded Spill ; RV64SFBSIFIVEU74-NEXT: sd s1, 8(sp) # 8-byte Folded Spill ; RV64SFBSIFIVEU74-NEXT: mv s1, a1 ; RV64SFBSIFIVEU74-NEXT: andi a0, a0, 1 +; RV64SFBSIFIVEU74-NEXT: sd s0, 16(sp) # 8-byte Folded Spill +; RV64SFBSIFIVEU74-NEXT: sd ra, 24(sp) # 8-byte Folded Spill ; RV64SFBSIFIVEU74-NEXT: mv s0, a2 ; RV64SFBSIFIVEU74-NEXT: beqz a0, .LBB16_4 ; RV64SFBSIFIVEU74-NEXT: # %bb.3: # %bb @@ -967,11 +967,11 @@ define void @sextw_removal_ccaddw(i1 %c, i32 signext %arg, i32 signext %arg1, i3 ; RV64SFBANDESAX45-LABEL: sextw_removal_ccaddw: ; RV64SFBANDESAX45: # %bb.0: # %bb ; RV64SFBANDESAX45-NEXT: addi sp, sp, -32 -; RV64SFBANDESAX45-NEXT: sd ra, 24(sp) # 8-byte Folded Spill ; RV64SFBANDESAX45-NEXT: sd s0, 16(sp) # 8-byte Folded Spill -; RV64SFBANDESAX45-NEXT: sd s1, 8(sp) # 8-byte Folded Spill ; RV64SFBANDESAX45-NEXT: mv s0, a2 +; RV64SFBANDESAX45-NEXT: sd s1, 8(sp) # 8-byte Folded Spill ; RV64SFBANDESAX45-NEXT: mv s1, a1 +; RV64SFBANDESAX45-NEXT: sd ra, 24(sp) # 8-byte Folded Spill ; RV64SFBANDESAX45-NEXT: nds.bbc a0, 0, .LBB16_2 ; RV64SFBANDESAX45-NEXT: # %bb.1: ; RV64SFBANDESAX45-NEXT: addw s1, s1, a3 @@ -991,11 +991,11 @@ define void @sextw_removal_ccaddw(i1 %c, i32 signext %arg, i32 signext %arg1, i3 ; ZICOND-LABEL: sextw_removal_ccaddw: ; ZICOND: # %bb.0: # %bb ; ZICOND-NEXT: addi sp, sp, -32 -; ZICOND-NEXT: sd ra, 24(sp) # 8-byte Folded Spill -; ZICOND-NEXT: sd s0, 16(sp) # 8-byte Folded Spill ; ZICOND-NEXT: sd s1, 8(sp) # 8-byte Folded Spill ; ZICOND-NEXT: mv s1, a1 ; ZICOND-NEXT: andi a0, a0, 1 +; ZICOND-NEXT: sd s0, 16(sp) # 8-byte Folded Spill +; ZICOND-NEXT: sd ra, 24(sp) # 8-byte Folded Spill ; ZICOND-NEXT: mv s0, a2 ; ZICOND-NEXT: beqz a0, .LBB16_4 ; ZICOND-NEXT: # %bb.3: # %bb @@ -1017,11 +1017,11 @@ define void @sextw_removal_ccaddw(i1 %c, i32 signext %arg, i32 signext %arg1, i3 ; RV32SFB-LABEL: sextw_removal_ccaddw: ; RV32SFB: # %bb.0: # %bb ; RV32SFB-NEXT: addi sp, sp, -16 -; RV32SFB-NEXT: sw ra, 12(sp) # 4-byte Folded Spill -; RV32SFB-NEXT: sw s0, 8(sp) # 4-byte Folded Spill ; RV32SFB-NEXT: sw s1, 4(sp) # 4-byte Folded Spill ; RV32SFB-NEXT: mv s1, a1 ; RV32SFB-NEXT: andi a0, a0, 1 +; RV32SFB-NEXT: sw s0, 8(sp) # 4-byte Folded Spill +; RV32SFB-NEXT: sw ra, 12(sp) # 4-byte Folded Spill ; RV32SFB-NEXT: mv s0, a2 ; RV32SFB-NEXT: beqz a0, .LBB16_4 ; RV32SFB-NEXT: # %bb.3: # %bb diff --git a/llvm/test/CodeGen/RISCV/zdinx-spill.ll b/llvm/test/CodeGen/RISCV/zdinx-spill.ll index d7a7006..6f206fe 100644 --- a/llvm/test/CodeGen/RISCV/zdinx-spill.ll +++ b/llvm/test/CodeGen/RISCV/zdinx-spill.ll @@ -9,7 +9,6 @@ define double @foo(double %x) nounwind { ; CHECK-NEXT: liveins: $x10, $x11, $x8, $x9, $x18, $x19, $x20, $x21, $x22, $x23, $x24, $x25, $x26, $x27 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: $x2 = frame-setup ADDI $x2, -64 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION def_cfa_offset 64 ; CHECK-NEXT: frame-setup SW killed $x8, $x2, 60 :: (store (s32) into %stack.1) ; CHECK-NEXT: frame-setup SW killed $x9, $x2, 56 :: (store (s32) into %stack.2) ; CHECK-NEXT: frame-setup SW killed $x18, $x2, 52 :: (store (s32) into %stack.3) @@ -22,18 +21,6 @@ define double @foo(double %x) nounwind { ; CHECK-NEXT: frame-setup SW killed $x25, $x2, 24 :: (store (s32) into %stack.10) ; CHECK-NEXT: frame-setup SW killed $x26, $x2, 20 :: (store (s32) into %stack.11) ; CHECK-NEXT: frame-setup SW killed $x27, $x2, 16 :: (store (s32) into %stack.12) - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x8, -4 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x9, -8 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x18, -12 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x19, -16 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x20, -20 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x21, -24 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x22, -28 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x23, -32 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x24, -36 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x25, -40 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x26, -44 - ; CHECK-NEXT: frame-setup CFI_INSTRUCTION offset $x27, -48 ; CHECK-NEXT: renamable $x10_x11 = nofpexcept FADD_D_IN32X killed renamable $x10_x11, renamable $x10_x11, 7, implicit $frm ; CHECK-NEXT: PseudoRV32ZdinxSD killed renamable $x10_x11, $x2, 8 :: (store (s64) into %stack.0, align 4) ; CHECK-NEXT: INLINEASM &"", 1 /* sideeffect attdialect */, 12 /* clobber */, implicit-def dead early-clobber $x6, 12 /* clobber */, implicit-def dead early-clobber $x7, 12 /* clobber */, implicit-def dead early-clobber $x8, 12 /* clobber */, implicit-def dead early-clobber $x9, 12 /* clobber */, implicit-def dead early-clobber $x10, 12 /* clobber */, implicit-def dead early-clobber $x11, 12 /* clobber */, implicit-def dead early-clobber $x12, 12 /* clobber */, implicit-def dead early-clobber $x13, 12 /* clobber */, implicit-def dead early-clobber $x14, 12 /* clobber */, implicit-def dead early-clobber $x15, 12 /* clobber */, implicit-def dead early-clobber $x16, 12 /* clobber */, implicit-def dead early-clobber $x17, 12 /* clobber */, implicit-def dead early-clobber $x18, 12 /* clobber */, implicit-def dead early-clobber $x19, 12 /* clobber */, implicit-def dead early-clobber $x20, 12 /* clobber */, implicit-def dead early-clobber $x21, 12 /* clobber */, implicit-def dead early-clobber $x22, 12 /* clobber */, implicit-def dead early-clobber $x23, 12 /* clobber */, implicit-def dead early-clobber $x24, 12 /* clobber */, implicit-def dead early-clobber $x25, 12 /* clobber */, implicit-def dead early-clobber $x26, 12 /* clobber */, implicit-def dead early-clobber $x27, 12 /* clobber */, implicit-def dead early-clobber $x28, 12 /* clobber */, implicit-def dead early-clobber $x29, 12 /* clobber */, implicit-def dead early-clobber $x31 @@ -50,20 +37,7 @@ define double @foo(double %x) nounwind { ; CHECK-NEXT: $x25 = frame-destroy LW $x2, 24 :: (load (s32) from %stack.10) ; CHECK-NEXT: $x26 = frame-destroy LW $x2, 20 :: (load (s32) from %stack.11) ; CHECK-NEXT: $x27 = frame-destroy LW $x2, 16 :: (load (s32) from %stack.12) - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x8 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x9 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x18 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x19 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x20 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x21 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x22 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x23 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x24 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x25 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x26 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION restore $x27 ; CHECK-NEXT: $x2 = frame-destroy ADDI $x2, 64 - ; CHECK-NEXT: frame-destroy CFI_INSTRUCTION def_cfa_offset 0 ; CHECK-NEXT: PseudoRET implicit $x10, implicit $x11 %a = fadd double %x, %x call void asm sideeffect "", "~{x6},~{x7},~{x8},~{x9},~{x10},~{x11},~{x12},~{x13},~{x14},~{x15},~{x16},~{x17},~{x18},~{x19},~{x20},~{x21},~{x22},~{x23},~{x24},~{x25},~{x26},~{x27},~{x28},~{x29},~{xr0},~{x31}"() diff --git a/llvm/test/CodeGen/X86/vector-bitreverse.ll b/llvm/test/CodeGen/X86/vector-bitreverse.ll index 5dcf190..834dfd6 100644 --- a/llvm/test/CodeGen/X86/vector-bitreverse.ll +++ b/llvm/test/CodeGen/X86/vector-bitreverse.ll @@ -8,7 +8,8 @@ ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512bw | FileCheck %s --check-prefixes=ALL,AVX,AVX512,AVX512BW ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+xop,+avx | FileCheck %s --check-prefixes=ALL,XOP,XOPAVX1 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+xop,+avx2 | FileCheck %s --check-prefixes=ALL,XOP,XOPAVX2 -; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+ssse3,+gfni | FileCheck %s --check-prefixes=ALL,GFNISSE +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+gfni | FileCheck %s --check-prefixes=ALL,GFNISSE,GFNISSE2 +; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+ssse3,+gfni | FileCheck %s --check-prefixes=ALL,GFNISSE,GFNISSSE3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx,+gfni | FileCheck %s --check-prefixes=ALL,GFNIAVX,GFNIAVX1 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+gfni | FileCheck %s --check-prefixes=ALL,GFNIAVX,GFNIAVX2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+gfni | FileCheck %s --check-prefixes=ALL,GFNIAVX,GFNIAVX512,GFNIAVX512F @@ -492,11 +493,20 @@ define <8 x i16> @test_bitreverse_v8i16(<8 x i16> %a) nounwind { ; XOP-NEXT: vpperm {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0, %xmm0 ; XOP-NEXT: retq ; -; GFNISSE-LABEL: test_bitreverse_v8i16: -; GFNISSE: # %bb.0: -; GFNISSE-NEXT: pshufb {{.*#+}} xmm0 = xmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14] -; GFNISSE-NEXT: gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 -; GFNISSE-NEXT: retq +; GFNISSE2-LABEL: test_bitreverse_v8i16: +; GFNISSE2: # %bb.0: +; GFNISSE2-NEXT: movdqa %xmm0, %xmm1 +; GFNISSE2-NEXT: psrlw $8, %xmm1 +; GFNISSE2-NEXT: psllw $8, %xmm0 +; GFNISSE2-NEXT: por %xmm1, %xmm0 +; GFNISSE2-NEXT: gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 +; GFNISSE2-NEXT: retq +; +; GFNISSSE3-LABEL: test_bitreverse_v8i16: +; GFNISSSE3: # %bb.0: +; GFNISSSE3-NEXT: pshufb {{.*#+}} xmm0 = xmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14] +; GFNISSSE3-NEXT: gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 +; GFNISSSE3-NEXT: retq ; ; GFNIAVX-LABEL: test_bitreverse_v8i16: ; GFNIAVX: # %bb.0: @@ -605,11 +615,25 @@ define <4 x i32> @test_bitreverse_v4i32(<4 x i32> %a) nounwind { ; XOP-NEXT: vpperm {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0, %xmm0 ; XOP-NEXT: retq ; -; GFNISSE-LABEL: test_bitreverse_v4i32: -; GFNISSE: # %bb.0: -; GFNISSE-NEXT: pshufb {{.*#+}} xmm0 = xmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12] -; GFNISSE-NEXT: gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 -; GFNISSE-NEXT: retq +; GFNISSE2-LABEL: test_bitreverse_v4i32: +; GFNISSE2: # %bb.0: +; GFNISSE2-NEXT: pxor %xmm1, %xmm1 +; GFNISSE2-NEXT: movdqa %xmm0, %xmm2 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm2, %xmm0 +; GFNISSE2-NEXT: gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 +; GFNISSE2-NEXT: retq +; +; GFNISSSE3-LABEL: test_bitreverse_v4i32: +; GFNISSSE3: # %bb.0: +; GFNISSSE3-NEXT: pshufb {{.*#+}} xmm0 = xmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12] +; GFNISSSE3-NEXT: gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 +; GFNISSSE3-NEXT: retq ; ; GFNIAVX-LABEL: test_bitreverse_v4i32: ; GFNIAVX: # %bb.0: @@ -720,11 +744,27 @@ define <2 x i64> @test_bitreverse_v2i64(<2 x i64> %a) nounwind { ; XOP-NEXT: vpperm {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0, %xmm0 ; XOP-NEXT: retq ; -; GFNISSE-LABEL: test_bitreverse_v2i64: -; GFNISSE: # %bb.0: -; GFNISSE-NEXT: pshufb {{.*#+}} xmm0 = xmm0[7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8] -; GFNISSE-NEXT: gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 -; GFNISSE-NEXT: retq +; GFNISSE2-LABEL: test_bitreverse_v2i64: +; GFNISSE2: # %bb.0: +; GFNISSE2-NEXT: pxor %xmm1, %xmm1 +; GFNISSE2-NEXT: movdqa %xmm0, %xmm2 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm2, %xmm0 +; GFNISSE2-NEXT: gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 +; GFNISSE2-NEXT: retq +; +; GFNISSSE3-LABEL: test_bitreverse_v2i64: +; GFNISSSE3: # %bb.0: +; GFNISSSE3-NEXT: pshufb {{.*#+}} xmm0 = xmm0[7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8] +; GFNISSSE3-NEXT: gf2p8affineqb $0, {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 +; GFNISSSE3-NEXT: retq ; ; GFNIAVX-LABEL: test_bitreverse_v2i64: ; GFNIAVX: # %bb.0: @@ -1042,15 +1082,30 @@ define <16 x i16> @test_bitreverse_v16i16(<16 x i16> %a) nounwind { ; XOPAVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; XOPAVX2-NEXT: retq ; -; GFNISSE-LABEL: test_bitreverse_v16i16: -; GFNISSE: # %bb.0: -; GFNISSE-NEXT: movdqa {{.*#+}} xmm2 = [1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14] -; GFNISSE-NEXT: pshufb %xmm2, %xmm0 -; GFNISSE-NEXT: movdqa {{.*#+}} xmm3 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm3, %xmm0 -; GFNISSE-NEXT: pshufb %xmm2, %xmm1 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm3, %xmm1 -; GFNISSE-NEXT: retq +; GFNISSE2-LABEL: test_bitreverse_v16i16: +; GFNISSE2: # %bb.0: +; GFNISSE2-NEXT: movdqa %xmm0, %xmm2 +; GFNISSE2-NEXT: psrlw $8, %xmm2 +; GFNISSE2-NEXT: psllw $8, %xmm0 +; GFNISSE2-NEXT: por %xmm2, %xmm0 +; GFNISSE2-NEXT: movdqa {{.*#+}} xmm2 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm2, %xmm0 +; GFNISSE2-NEXT: movdqa %xmm1, %xmm3 +; GFNISSE2-NEXT: psrlw $8, %xmm3 +; GFNISSE2-NEXT: psllw $8, %xmm1 +; GFNISSE2-NEXT: por %xmm3, %xmm1 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm2, %xmm1 +; GFNISSE2-NEXT: retq +; +; GFNISSSE3-LABEL: test_bitreverse_v16i16: +; GFNISSSE3: # %bb.0: +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm2 = [1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14] +; GFNISSSE3-NEXT: pshufb %xmm2, %xmm0 +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm3 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm3, %xmm0 +; GFNISSSE3-NEXT: pshufb %xmm2, %xmm1 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm3, %xmm1 +; GFNISSSE3-NEXT: retq ; ; GFNIAVX1-LABEL: test_bitreverse_v16i16: ; GFNIAVX1: # %bb.0: @@ -1241,15 +1296,39 @@ define <8 x i32> @test_bitreverse_v8i32(<8 x i32> %a) nounwind { ; XOPAVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; XOPAVX2-NEXT: retq ; -; GFNISSE-LABEL: test_bitreverse_v8i32: -; GFNISSE: # %bb.0: -; GFNISSE-NEXT: movdqa {{.*#+}} xmm2 = [3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12] -; GFNISSE-NEXT: pshufb %xmm2, %xmm0 -; GFNISSE-NEXT: movdqa {{.*#+}} xmm3 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm3, %xmm0 -; GFNISSE-NEXT: pshufb %xmm2, %xmm1 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm3, %xmm1 -; GFNISSE-NEXT: retq +; GFNISSE2-LABEL: test_bitreverse_v8i32: +; GFNISSE2: # %bb.0: +; GFNISSE2-NEXT: pxor %xmm2, %xmm2 +; GFNISSE2-NEXT: movdqa %xmm0, %xmm3 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm3, %xmm0 +; GFNISSE2-NEXT: movdqa {{.*#+}} xmm3 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm3, %xmm0 +; GFNISSE2-NEXT: movdqa %xmm1, %xmm4 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm2[8],xmm4[9],xmm2[9],xmm4[10],xmm2[10],xmm4[11],xmm2[11],xmm4[12],xmm2[12],xmm4[13],xmm2[13],xmm4[14],xmm2[14],xmm4[15],xmm2[15] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm4, %xmm1 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm3, %xmm1 +; GFNISSE2-NEXT: retq +; +; GFNISSSE3-LABEL: test_bitreverse_v8i32: +; GFNISSSE3: # %bb.0: +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm2 = [3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12] +; GFNISSSE3-NEXT: pshufb %xmm2, %xmm0 +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm3 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm3, %xmm0 +; GFNISSSE3-NEXT: pshufb %xmm2, %xmm1 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm3, %xmm1 +; GFNISSSE3-NEXT: retq ; ; GFNIAVX1-LABEL: test_bitreverse_v8i32: ; GFNIAVX1: # %bb.0: @@ -1444,15 +1523,43 @@ define <4 x i64> @test_bitreverse_v4i64(<4 x i64> %a) nounwind { ; XOPAVX2-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 ; XOPAVX2-NEXT: retq ; -; GFNISSE-LABEL: test_bitreverse_v4i64: -; GFNISSE: # %bb.0: -; GFNISSE-NEXT: movdqa {{.*#+}} xmm2 = [7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8] -; GFNISSE-NEXT: pshufb %xmm2, %xmm0 -; GFNISSE-NEXT: movdqa {{.*#+}} xmm3 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm3, %xmm0 -; GFNISSE-NEXT: pshufb %xmm2, %xmm1 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm3, %xmm1 -; GFNISSE-NEXT: retq +; GFNISSE2-LABEL: test_bitreverse_v4i64: +; GFNISSE2: # %bb.0: +; GFNISSE2-NEXT: pxor %xmm2, %xmm2 +; GFNISSE2-NEXT: movdqa %xmm0, %xmm3 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm3, %xmm0 +; GFNISSE2-NEXT: movdqa {{.*#+}} xmm3 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm3, %xmm0 +; GFNISSE2-NEXT: movdqa %xmm1, %xmm4 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm2[8],xmm4[9],xmm2[9],xmm4[10],xmm2[10],xmm4[11],xmm2[11],xmm4[12],xmm2[12],xmm4[13],xmm2[13],xmm4[14],xmm2[14],xmm4[15],xmm2[15] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm4 = xmm4[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm4 = xmm4[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm4, %xmm1 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm3, %xmm1 +; GFNISSE2-NEXT: retq +; +; GFNISSSE3-LABEL: test_bitreverse_v4i64: +; GFNISSSE3: # %bb.0: +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm2 = [7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8] +; GFNISSSE3-NEXT: pshufb %xmm2, %xmm0 +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm3 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm3, %xmm0 +; GFNISSSE3-NEXT: pshufb %xmm2, %xmm1 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm3, %xmm1 +; GFNISSSE3-NEXT: retq ; ; GFNIAVX1-LABEL: test_bitreverse_v4i64: ; GFNIAVX1: # %bb.0: @@ -2035,19 +2142,44 @@ define <32 x i16> @test_bitreverse_v32i16(<32 x i16> %a) nounwind { ; XOPAVX2-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; XOPAVX2-NEXT: retq ; -; GFNISSE-LABEL: test_bitreverse_v32i16: -; GFNISSE: # %bb.0: -; GFNISSE-NEXT: movdqa {{.*#+}} xmm4 = [1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14] -; GFNISSE-NEXT: pshufb %xmm4, %xmm0 -; GFNISSE-NEXT: movdqa {{.*#+}} xmm5 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm0 -; GFNISSE-NEXT: pshufb %xmm4, %xmm1 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm1 -; GFNISSE-NEXT: pshufb %xmm4, %xmm2 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm2 -; GFNISSE-NEXT: pshufb %xmm4, %xmm3 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm3 -; GFNISSE-NEXT: retq +; GFNISSE2-LABEL: test_bitreverse_v32i16: +; GFNISSE2: # %bb.0: +; GFNISSE2-NEXT: movdqa %xmm0, %xmm4 +; GFNISSE2-NEXT: psrlw $8, %xmm4 +; GFNISSE2-NEXT: psllw $8, %xmm0 +; GFNISSE2-NEXT: por %xmm4, %xmm0 +; GFNISSE2-NEXT: movdqa {{.*#+}} xmm4 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm4, %xmm0 +; GFNISSE2-NEXT: movdqa %xmm1, %xmm5 +; GFNISSE2-NEXT: psrlw $8, %xmm5 +; GFNISSE2-NEXT: psllw $8, %xmm1 +; GFNISSE2-NEXT: por %xmm5, %xmm1 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm4, %xmm1 +; GFNISSE2-NEXT: movdqa %xmm2, %xmm5 +; GFNISSE2-NEXT: psrlw $8, %xmm5 +; GFNISSE2-NEXT: psllw $8, %xmm2 +; GFNISSE2-NEXT: por %xmm5, %xmm2 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm4, %xmm2 +; GFNISSE2-NEXT: movdqa %xmm3, %xmm5 +; GFNISSE2-NEXT: psrlw $8, %xmm5 +; GFNISSE2-NEXT: psllw $8, %xmm3 +; GFNISSE2-NEXT: por %xmm5, %xmm3 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm4, %xmm3 +; GFNISSE2-NEXT: retq +; +; GFNISSSE3-LABEL: test_bitreverse_v32i16: +; GFNISSSE3: # %bb.0: +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm4 = [1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14] +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm0 +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm5 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm0 +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm1 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm1 +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm2 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm2 +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm3 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm3 +; GFNISSSE3-NEXT: retq ; ; GFNIAVX1-LABEL: test_bitreverse_v32i16: ; GFNIAVX1: # %bb.0: @@ -2393,19 +2525,61 @@ define <16 x i32> @test_bitreverse_v16i32(<16 x i32> %a) nounwind { ; XOPAVX2-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; XOPAVX2-NEXT: retq ; -; GFNISSE-LABEL: test_bitreverse_v16i32: -; GFNISSE: # %bb.0: -; GFNISSE-NEXT: movdqa {{.*#+}} xmm4 = [3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12] -; GFNISSE-NEXT: pshufb %xmm4, %xmm0 -; GFNISSE-NEXT: movdqa {{.*#+}} xmm5 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm0 -; GFNISSE-NEXT: pshufb %xmm4, %xmm1 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm1 -; GFNISSE-NEXT: pshufb %xmm4, %xmm2 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm2 -; GFNISSE-NEXT: pshufb %xmm4, %xmm3 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm3 -; GFNISSE-NEXT: retq +; GFNISSE2-LABEL: test_bitreverse_v16i32: +; GFNISSE2: # %bb.0: +; GFNISSE2-NEXT: pxor %xmm4, %xmm4 +; GFNISSE2-NEXT: movdqa %xmm0, %xmm5 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm5 = xmm5[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3],xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm5, %xmm0 +; GFNISSE2-NEXT: movdqa {{.*#+}} xmm5 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm5, %xmm0 +; GFNISSE2-NEXT: movdqa %xmm1, %xmm6 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm6 = xmm6[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm6, %xmm1 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm5, %xmm1 +; GFNISSE2-NEXT: movdqa %xmm2, %xmm6 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm6 = xmm6[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm6, %xmm2 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm5, %xmm2 +; GFNISSE2-NEXT: movdqa %xmm3, %xmm6 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm6 = xmm6[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm6, %xmm3 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm5, %xmm3 +; GFNISSE2-NEXT: retq +; +; GFNISSSE3-LABEL: test_bitreverse_v16i32: +; GFNISSSE3: # %bb.0: +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm4 = [3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12] +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm0 +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm5 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm0 +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm1 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm1 +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm2 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm2 +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm3 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm3 +; GFNISSSE3-NEXT: retq ; ; GFNIAVX1-LABEL: test_bitreverse_v16i32: ; GFNIAVX1: # %bb.0: @@ -2759,19 +2933,69 @@ define <8 x i64> @test_bitreverse_v8i64(<8 x i64> %a) nounwind { ; XOPAVX2-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1 ; XOPAVX2-NEXT: retq ; -; GFNISSE-LABEL: test_bitreverse_v8i64: -; GFNISSE: # %bb.0: -; GFNISSE-NEXT: movdqa {{.*#+}} xmm4 = [7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8] -; GFNISSE-NEXT: pshufb %xmm4, %xmm0 -; GFNISSE-NEXT: movdqa {{.*#+}} xmm5 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm0 -; GFNISSE-NEXT: pshufb %xmm4, %xmm1 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm1 -; GFNISSE-NEXT: pshufb %xmm4, %xmm2 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm2 -; GFNISSE-NEXT: pshufb %xmm4, %xmm3 -; GFNISSE-NEXT: gf2p8affineqb $0, %xmm5, %xmm3 -; GFNISSE-NEXT: retq +; GFNISSE2-LABEL: test_bitreverse_v8i64: +; GFNISSE2: # %bb.0: +; GFNISSE2-NEXT: pxor %xmm4, %xmm4 +; GFNISSE2-NEXT: movdqa %xmm0, %xmm5 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm4[8],xmm5[9],xmm4[9],xmm5[10],xmm4[10],xmm5[11],xmm4[11],xmm5[12],xmm4[12],xmm5[13],xmm4[13],xmm5[14],xmm4[14],xmm5[15],xmm4[15] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm5[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm5 = xmm5[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm5 = xmm5[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3],xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm5, %xmm0 +; GFNISSE2-NEXT: movdqa {{.*#+}} xmm5 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm5, %xmm0 +; GFNISSE2-NEXT: movdqa %xmm1, %xmm6 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm6[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm6 = xmm6[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3],xmm1[4],xmm4[4],xmm1[5],xmm4[5],xmm1[6],xmm4[6],xmm1[7],xmm4[7] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm6, %xmm1 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm5, %xmm1 +; GFNISSE2-NEXT: movdqa %xmm2, %xmm6 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm6[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm6 = xmm6[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3],xmm2[4],xmm4[4],xmm2[5],xmm4[5],xmm2[6],xmm4[6],xmm2[7],xmm4[7] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm6, %xmm2 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm5, %xmm2 +; GFNISSE2-NEXT: movdqa %xmm3, %xmm6 +; GFNISSE2-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm4[8],xmm6[9],xmm4[9],xmm6[10],xmm4[10],xmm6[11],xmm4[11],xmm6[12],xmm4[12],xmm6[13],xmm4[13],xmm6[14],xmm4[14],xmm6[15],xmm4[15] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm6[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm6 = xmm6[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm6 = xmm6[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3],xmm3[4],xmm4[4],xmm3[5],xmm4[5],xmm3[6],xmm4[6],xmm3[7],xmm4[7] +; GFNISSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[2,3,0,1] +; GFNISSE2-NEXT: pshuflw {{.*#+}} xmm3 = xmm3[3,2,1,0,4,5,6,7] +; GFNISSE2-NEXT: pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,7,6,5,4] +; GFNISSE2-NEXT: packuswb %xmm6, %xmm3 +; GFNISSE2-NEXT: gf2p8affineqb $0, %xmm5, %xmm3 +; GFNISSE2-NEXT: retq +; +; GFNISSSE3-LABEL: test_bitreverse_v8i64: +; GFNISSSE3: # %bb.0: +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm4 = [7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8] +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm0 +; GFNISSSE3-NEXT: movdqa {{.*#+}} xmm5 = [1,2,4,8,16,32,64,128,1,2,4,8,16,32,64,128] +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm0 +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm1 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm1 +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm2 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm2 +; GFNISSSE3-NEXT: pshufb %xmm4, %xmm3 +; GFNISSSE3-NEXT: gf2p8affineqb $0, %xmm5, %xmm3 +; GFNISSSE3-NEXT: retq ; ; GFNIAVX1-LABEL: test_bitreverse_v8i64: ; GFNIAVX1: # %bb.0: |