aboutsummaryrefslogtreecommitdiff
path: root/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll')
-rw-r--r--llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll2946
1 files changed, 1516 insertions, 1430 deletions
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: