1 files changed, 541 insertions, 534 deletions
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]