[llvm] [AMDGPU][True16][CodeGen] use vgpr16 for zext patterns (PR #153894)
Brox Chen via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 15 15:45:56 PDT 2025
https://github.com/broxigarchen created https://github.com/llvm/llvm-project/pull/153894
Update true16 mode with zext patterns using vgpr16 for 16bit data types
>From f26b3846624fd495d105f800cff9512cd9432f9b Mon Sep 17 00:00:00 2001
From: guochen2 <guochen2 at amd.com>
Date: Fri, 15 Aug 2025 17:57:21 -0400
Subject: [PATCH] fix zext
---
llvm/lib/Target/AMDGPU/SIInstructions.td | 22 +
llvm/test/CodeGen/AMDGPU/GlobalISel/mul.ll | 2 +-
.../CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll | 11901 ++++++++--------
.../CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll | 1148 +-
.../CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll | 1320 +-
.../CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll | 2886 ++--
.../CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll | 240 +-
.../CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll | 5414 ++++---
.../CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll | 637 +-
.../CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll | 594 +-
.../AMDGPU/amdgpu-llvm-debuginfo-analyzer.ll | 1 +
.../atomic_optimizations_global_pointer.ll | 64 +-
llvm/test/CodeGen/AMDGPU/bf16.ll | 14 +-
.../buffer-fat-pointer-atomicrmw-fadd.ll | 42 +-
.../buffer-fat-pointer-atomicrmw-fmax.ll | 42 +-
.../buffer-fat-pointer-atomicrmw-fmin.ll | 42 +-
.../CodeGen/AMDGPU/calling-conventions.ll | 100 +-
llvm/test/CodeGen/AMDGPU/clamp-modifier.ll | 4 +-
llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll | 42 +-
.../test/CodeGen/AMDGPU/dynamic_stackalloc.ll | 5 +-
.../CodeGen/AMDGPU/flat-atomicrmw-fadd.ll | 106 +-
.../CodeGen/AMDGPU/flat-atomicrmw-fmax.ll | 110 +-
.../CodeGen/AMDGPU/flat-atomicrmw-fmin.ll | 110 +-
.../CodeGen/AMDGPU/flat-atomicrmw-fsub.ll | 106 +-
llvm/test/CodeGen/AMDGPU/fmul-to-ldexp.ll | 2 +-
.../AMDGPU/fold-int-pow2-with-fmul-or-fdiv.ll | 6 +-
llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll | 6 +-
llvm/test/CodeGen/AMDGPU/function-args.ll | 251 +-
.../AMDGPU/gfx-callable-argument-types.ll | 222 +-
.../CodeGen/AMDGPU/global-atomicrmw-fadd.ll | 106 +-
.../CodeGen/AMDGPU/global-atomicrmw-fmax.ll | 110 +-
.../CodeGen/AMDGPU/global-atomicrmw-fmin.ll | 110 +-
.../CodeGen/AMDGPU/global-atomicrmw-fsub.ll | 106 +-
llvm/test/CodeGen/AMDGPU/idot4u.ll | 41 +-
.../CodeGen/AMDGPU/integer-mad-patterns.ll | 28 +-
.../CodeGen/AMDGPU/local-atomicrmw-fadd.ll | 60 +-
.../CodeGen/AMDGPU/local-atomicrmw-fmax.ll | 68 +-
.../CodeGen/AMDGPU/local-atomicrmw-fmin.ll | 68 +-
.../CodeGen/AMDGPU/local-atomicrmw-fsub.ll | 60 +-
llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll | 31 +-
llvm/test/CodeGen/AMDGPU/mad.u16.ll | 7 +-
llvm/test/CodeGen/AMDGPU/preserve-hi16.ll | 54 +-
.../CodeGen/AMDGPU/shrink-add-sub-constant.ll | 6 +-
llvm/test/CodeGen/AMDGPU/vector-reduce-add.ll | 126 +-
.../test/CodeGen/AMDGPU/vector-reduce-umin.ll | 78 +-
45 files changed, 12480 insertions(+), 14018 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index bd5dfa92a8e43..6488fa3dacfb3 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -3056,6 +3056,8 @@ def : GCNPat<
}
} // AddedComplexity = 1
+foreach p = [NotHasTrue16BitInsts, UseFakeTrue16Insts] in
+let True16Predicate = p in {
def : GCNPat<
(i32 (DivergentUnaryFrag<zext> i16:$src)),
(V_AND_B32_e64 (S_MOV_B32 (i32 0xffff)), $src)
@@ -3071,6 +3073,26 @@ def : GCNPat<
def : GCNPat<
(i32 (zext (i16 (bitconvert fp16_zeros_high_16bits:$src)))),
(COPY VSrc_b16:$src)>;
+}
+
+let True16Predicate = UseRealTrue16Insts in {
+def : GCNPat<
+ (i32 (DivergentUnaryFrag<zext> i16:$src)),
+ (REG_SEQUENCE VGPR_32, $src, lo16, (V_MOV_B16_t16_e64 0, (i16 0), 0), hi16)
+>;
+
+def : GCNPat<
+ (i64 (DivergentUnaryFrag<zext> i16:$src)),
+ (REG_SEQUENCE VReg_64,
+ (REG_SEQUENCE VGPR_32, $src, lo16, (V_MOV_B16_t16_e64 0, (i16 0), 0), hi16), sub0,
+ (S_MOV_B32 (i32 0)), sub1)
+>;
+
+def : GCNPat<
+ (i32 (zext (i16 (bitconvert fp16_zeros_high_16bits:$src)))),
+ (REG_SEQUENCE VGPR_32, $src, lo16, (V_MOV_B16_t16_e64 0, (i16 0), 0), hi16)
+>;
+}
def : GCNPat <
(i32 (trunc i64:$a)),
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/mul.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/mul.ll
index 01854c8560ce2..637aaf7529364 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/mul.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/mul.ll
@@ -164,7 +164,7 @@ define zeroext i16 @v_mul_i16_zeroext(i16 zeroext %num, i16 zeroext %den) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mul_lo_u16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_mul_i16_zeroext:
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
index 0d5f538215f18..d03d6a8940b2f 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll
@@ -6309,64 +6309,64 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2)
@@ -6394,50 +6394,50 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
@@ -6498,50 +6498,50 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT: .LBB12_4: ; %end
@@ -6549,307 +6549,266 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v1.h, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff, v66
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v55, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v150.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v66, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v55, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v4.l, v4.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v55, v39
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v149.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l
+; 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_or_b32_e32 v4, v39, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v148.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v147.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v55, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v55, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v135.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v8.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v133.l
; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l
; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v132.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v53, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v131.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v53, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v119.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v116.l
; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v129.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v114.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v113.l
; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v53
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v51.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v53
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v118.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v51, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v102.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v51, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v116.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v115.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v99.l
; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v50, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v114.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v50, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v113.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v96.l
; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l
; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v49, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v112.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l
+; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v85.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l
; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v50
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v18.l, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v50
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v49, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v19.l, v19.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v20.l
-; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v101.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v48, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v49
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v82.l
; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v81.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v71.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v48, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v22.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v49
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v21.l, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v98.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l
; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v22.l, v22.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v97.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v69.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v87.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v24.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v25.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v84.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v27.l, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v81.l
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v28.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v80.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v71.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v70.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v69.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v30.l, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v39
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v31.l, v31.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l
-; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v68.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v32.l, v32.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v39
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32
; GFX11-TRUE16-NEXT: s_clause 0x5
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
@@ -15413,63 +15372,63 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:384
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:372
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:368
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:364
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:360
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:356
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:348
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:344
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:360
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:352
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:344
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:324
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:316
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:264
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:288
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:280
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:272
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:264
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260
; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:256
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:256
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:240
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:240
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v103, off, s32 offset:388
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:232
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:228
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:224
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:220
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:216
+; GFX11-TRUE16-NEXT: scratch_load_b32 v114, off, s32 offset:388
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:16
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:16
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:80
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:32
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:40
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:48
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:56
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:64
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:72
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:80
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:96
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:96
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:112
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:112
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v161, off, s32 offset:136
@@ -15483,144 +15442,143 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:212
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:204
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:196
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:188
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:180
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:172
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:164
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:140
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:132
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:116
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:100
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:84
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:52
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:44
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:20
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v116, off, s32 offset:4
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.h, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v12.l
; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.l, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v29.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v29.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54)
-; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v103
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(53)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v80.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v81.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(52)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v82.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v82.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v82.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v83.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(50)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v83.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v83.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(49)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v84.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v85.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(48)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.h, 8, v85.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v85.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v86.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(46)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v87.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v87.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(45)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v96.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v97.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(44)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.h, 8, v97.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v97.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v98.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(42)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v99.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v99.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(41)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v99.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(40)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v101.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v101.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v102.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(38)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v160.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(37)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v160.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(36)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v161.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.h, 8, v161.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v161.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v161.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(34)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v162.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v162.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v162.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v162.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(32)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v163.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v163.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(30)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v164.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v164.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v164.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(28)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v165.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v165.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v70.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v70.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v68.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v65.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v165.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v71.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v71.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v70.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v67.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v66.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v64.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.l, 8, v55.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v54.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v54.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v53.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v52.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v51.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v49.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v54.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v53.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v52.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v51.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v50.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v31.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
@@ -15634,746 +15592,660 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB14_3: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v151.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v148.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v150.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v146.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v151.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v150.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v147.h
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v0.h, v149.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v149.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v144.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v133.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v148.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v144.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v145.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v1.h, v147.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v146.l
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v0.l, v151.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v151.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v0.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v150.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v149, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v1.h, v150.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v132.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v131.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v148.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v130.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v149, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v2.l, v148.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v149.h
; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v131.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v2.l, v146.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v131.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v135.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v3.l, v145.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v128.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v118.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v134.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v145.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v130.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v149, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v3.l, v147.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v147.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v118.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v149, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v4.l, v144.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v4.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v133.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v117.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v116.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v4.l, v135.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v134.l
-; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v5.l, v133.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v132.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v114.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v6.l, v132.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v130.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v7.l, v130.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v129.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v102.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v8.l, v128.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v119.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v9.l, v118.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v102.h
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v117.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v97.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v10.l, v116.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v100.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v115.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v11.l, v114.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v11.h, v113.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v85.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v12.l, v113.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v103.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v13.l, v103.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v101.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v81.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v14.l, v101.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v14.h, v99.h
-; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v71.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v15.l, v99.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v15.h, v98.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v16.l, v97.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v71.h
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v16.h, v96.h
-; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v67.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v17.l, v87.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v149, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v5.l, v135.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v5.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v132.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v129.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v113.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v112.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v149, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v6.l, v133.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v6.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v128.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v103.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v101.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v149, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v7.l, v131.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v7.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v100.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v99.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v149, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v8.l, v129.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v116.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v114.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v96.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v149, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v9.l, v128.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v9.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v113.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v86.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v149, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v10.l, v117.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v10.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v102.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v84.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v82.h
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v81.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v149, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v11.l, v116.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v11.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.l
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v99.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v149, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v12.l, v114.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v12.l, v149.h
; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v17.h, v86.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v64.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v18.l, v85.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v19.l
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v67.h
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v18.h, v84.h
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v51.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v19.l, v83.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v97.h
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v69.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v68.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v149, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v13.l, v112.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v13.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v87.h
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v67.l
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v19.h, v82.h
-; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v48.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v20.l, v82.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v20.h, v80.h
-; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v39.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v21.l, v80.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v21.h, v70.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v22.l, v70.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v22.h, v68.h
-; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v38.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v23.l, v68.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v38.h
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v23.h, v66.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v36.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v24.l, v66.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v37.h
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v24.h, v65.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v36.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v27, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v25.l, v64.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v25.h, v55.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v28, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v26.l, v55.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v35.h
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v26.h, v54.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v27.l, v54.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v65.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v149, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v14.l, v102.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v14.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v85.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v83.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v55.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v50.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v149, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v15.l, v100.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v15.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v82.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v149, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v16.l, v98.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v16.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.h
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v39.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v149, v16
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v17.l, v97.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v17.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v70.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v68.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v38.h
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v38.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v149, v17
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v18.l, v87.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v18.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v66.h
+; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v149, v18
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v19.l, v85.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v19.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v149, v19
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v20.l, v83.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v20.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v55.l
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v54.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v34.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v149, v20
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v21.l, v81.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v21.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v33.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v149, v21
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v22.l, v71.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v22.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v32.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v28.h, v53.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v28.l, v53.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v33.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v33.h
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v28.h, v52.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v33.l, v29.h, v51.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v29.l, v50.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v149, v22
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v23.l, v70.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v23.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v51.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v149, v23
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v24.l, v67.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v24.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v149, v24
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v25.l, v66.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v25.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v149, v25
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v26.l, v64.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v26.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v149, v26
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v27.l, v54.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v27.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v149, v27
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v28.l, v53.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v28.l, v149.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v149, v28
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v29.l, v52.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v29.l, v149.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v30.l, v50.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v32.l, v49.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v149, v29
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v30.l, v51.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v30.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v149, v30
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v31.l, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v31.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v149, v31
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_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, v151.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v148.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v150.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v146.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v147.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v146.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, 0
; 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 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v149.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v151.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v150.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v134.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v31, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v148.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v149.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v147.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v134.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v148.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v145.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v31, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v31.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v145.h, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v131.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v31, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v130.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v130.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v144.h, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v145.l, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v135.h, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v135.l, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.l, v32.l, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v134.l, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v118.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v133.l, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v31, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v31, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v118.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v117.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v133.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v133.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v132.h, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v31, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v131.h, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v132.l, v6.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v132.l, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v117.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v114.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v31, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v113.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v129.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v129.h, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.l, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v112.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v128.l, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v119.h, v9.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v102.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v118.l, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v102.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v31, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v128.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v128.h, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v103.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v31, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v101.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v117.h, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v118.l, v9.h
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v117.h, v10.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v100.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v11.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v116.l, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v100.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v115.h, v11.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v97.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v114.h, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v31, v12
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v116.l, v10.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v116.h, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v99.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v31, v13
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v96.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v96.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v87.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v96.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v103.h, v13.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v85.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v103.l, v13.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v13.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v31, v14
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.h, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v113.l, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v86.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v31, v15
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v84.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v84.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v102.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v102.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v101.h, v14.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v83.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v15.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v101.l, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v84.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v99.h, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v31, v16
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v100.h, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.l, v14.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v82.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v99.l, v15.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v15.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v31, v17
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v80.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v80.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v98.h, v15.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v99.l, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v98.l, v16.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v71.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v97.l, v16.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v71.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v16.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v16.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v87.l, v17.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v31, v18
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v97.l, v16.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v97.h, v16.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v17.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v31, v19
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v68.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v67.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v87.l, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v87.h, v17.h
; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v86.h, v18.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v67.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v85.l, v18.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v67.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v18.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v84.h, v19.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.h, v19.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v31, v20
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v85.l, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v85.h, v18.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v19.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v65.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v31, v21
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v55.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v50.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v83.l, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v83.h, v19.h
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v82.h, v20.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v51.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v21.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v82.l, v20.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v52.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v20.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v80.h, v21.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v22.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v80.l, v21.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v21.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v31, v22
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v81.h, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v82.l, v20.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v31, v23
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v48.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.l, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v71.h, v21.h
; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v70.h, v22.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v39.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.l, v22.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v22.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v22.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v68.h, v23.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v68.l, v23.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v23.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v31, v24
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v70.l, v22.l
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v70.h, v22.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v31, v25
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v38.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v67.h, v23.l
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v68.l, v23.h
; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v37.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v66.h, v24.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v64.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v26, v31
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v65.l, v25.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v25.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v26
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v25.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.l, v24.l
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v66.h, v24.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v27
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v25.h
+; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v36.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v64.l, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v64.h, v25.h
; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l
; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v27, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v55.l, v26.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v55.h, v26.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v35.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v28
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v27.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v28.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v53.h, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v31, v28
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v27.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v26.l
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v55.l, v26.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v31, v29
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v34.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v53.h, v27.l
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v54.l, v27.h
; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v33.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v28.h
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v31, v30
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v34.h, 0x300, v29.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v28.l
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v53.l, v28.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v53.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v30.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v31, v34
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v29.h
+; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v28.l
-; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v32.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v51.h, v29.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v52.l, v29.h
; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.l, v30.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v33
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v50.l, v30.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.l, 0x300, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v30.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v49.h, v30.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v32.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v31, v33
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v32.h
+; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v50.h, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v51.l, v30.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v31, v33
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.h, 0x300, v32.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v32
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -42156,64 +42028,64 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2)
@@ -42241,50 +42113,50 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
@@ -42328,50 +42200,50 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT: .LBB36_4: ; %end
@@ -42379,307 +42251,266 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v1.h, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff, v66
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v55, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v150.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v66, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v55, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v4.l, v4.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v55, v39
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v149.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l
+; 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_or_b32_e32 v4, v39, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v148.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v147.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v55, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v55, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v135.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v8.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v133.l
; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l
; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v132.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v53, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v131.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v53, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v119.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v116.l
; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v129.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v114.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v113.l
; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v53
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v51.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v53
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v118.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v51, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v102.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v51, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v116.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v115.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v99.l
; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v50, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v114.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v50, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v113.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v96.l
; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l
; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v49, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v112.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l
+; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v85.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l
; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v50
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v18.l, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v50
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v49, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v19.l, v19.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v20.l
-; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v101.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v48, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v49
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v82.l
; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v81.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v71.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v48, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v22.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v49
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v21.l, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v98.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l
; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v22.l, v22.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v97.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v69.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v87.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v24.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v25.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v84.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v27.l, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v81.l
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v28.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v80.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v71.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v70.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v69.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v30.l, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v39
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v31.l, v31.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l
-; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v68.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v32.l, v32.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v39
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32
; GFX11-TRUE16-NEXT: s_clause 0x5
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
@@ -52210,63 +52041,63 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:384
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:372
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:368
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:364
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:360
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:356
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:348
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:344
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:360
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:352
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:344
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:324
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:316
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:264
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:288
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:280
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:272
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:264
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260
; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:256
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:256
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:240
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:240
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v103, off, s32 offset:388
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:232
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:228
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:224
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:220
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:216
+; GFX11-TRUE16-NEXT: scratch_load_b32 v114, off, s32 offset:388
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:16
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:16
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:80
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:32
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:40
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:48
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:56
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:64
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:72
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:80
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:96
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:96
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:112
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:112
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v161, off, s32 offset:136
@@ -52280,144 +52111,143 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:212
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:204
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:196
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:188
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:180
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:172
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:164
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:140
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:132
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:116
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:100
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:84
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:52
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:44
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:20
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v116, off, s32 offset:4
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.h, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v12.l
; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.l, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v29.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v29.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54)
-; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v103
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(53)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v80.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v81.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(52)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v82.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v82.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v82.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v83.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(50)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v83.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v83.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(49)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v84.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v85.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(48)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.h, 8, v85.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v85.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v86.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(46)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v87.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v87.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(45)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v96.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v97.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(44)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.h, 8, v97.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v97.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v98.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(42)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v99.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v99.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(41)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v99.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(40)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v101.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v101.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v102.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(38)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v160.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(37)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v160.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(36)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v161.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.h, 8, v161.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v161.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v161.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(34)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v162.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v162.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v162.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v162.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(32)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v163.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v163.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(30)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v164.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v164.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v164.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(28)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v165.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v165.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v70.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v70.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v68.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v65.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v165.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v71.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v71.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v70.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v67.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v66.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v64.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.l, 8, v55.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v54.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v54.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v53.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v52.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v51.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v49.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v54.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v53.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v52.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v51.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v50.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v31.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
@@ -52431,746 +52261,660 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB38_3: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v151.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v148.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v150.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v146.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v151.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v150.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v147.h
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v0.h, v149.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v149.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v144.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v133.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v148.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v144.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v145.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v1.h, v147.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v146.l
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v0.l, v151.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v151.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v0.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v150.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v149, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v1.h, v150.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v132.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v131.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v148.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v130.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v149, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v2.l, v148.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v149.h
; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v131.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v2.l, v146.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v131.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v135.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v3.l, v145.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v128.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v118.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v134.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v145.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v130.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v149, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v3.l, v147.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v147.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v118.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v149, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v4.l, v144.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v4.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v133.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v117.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v116.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v4.l, v135.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v134.l
-; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v5.l, v133.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v132.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v114.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v6.l, v132.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v130.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v7.l, v130.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v129.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v102.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v8.l, v128.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v119.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v9.l, v118.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v102.h
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v117.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v97.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v10.l, v116.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v100.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v115.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v11.l, v114.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v11.h, v113.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v85.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v12.l, v113.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v103.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v13.l, v103.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v101.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v81.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v14.l, v101.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v14.h, v99.h
-; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v71.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v15.l, v99.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v15.h, v98.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v16.l, v97.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v71.h
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v16.h, v96.h
-; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v67.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v17.l, v87.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v149, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v5.l, v135.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v5.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v132.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v129.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v113.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v112.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v149, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v6.l, v133.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v6.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v128.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v103.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v101.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v149, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v7.l, v131.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v7.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v100.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v99.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v149, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v8.l, v129.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v116.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v114.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v96.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v149, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v9.l, v128.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v9.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v113.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v86.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v149, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v10.l, v117.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v10.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v102.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v84.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v82.h
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v81.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v149, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v11.l, v116.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v11.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.l
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v99.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v149, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v12.l, v114.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v12.l, v149.h
; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v17.h, v86.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v64.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v18.l, v85.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v19.l
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v67.h
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v18.h, v84.h
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v51.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v19.l, v83.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v97.h
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v69.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v68.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v149, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v13.l, v112.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v13.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v87.h
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v67.l
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v19.h, v82.h
-; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v48.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v20.l, v82.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v20.h, v80.h
-; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v39.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v21.l, v80.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v21.h, v70.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v22.l, v70.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v22.h, v68.h
-; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v38.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v23.l, v68.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v38.h
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v23.h, v66.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v36.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v24.l, v66.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v37.h
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v24.h, v65.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v36.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v27, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v25.l, v64.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v25.h, v55.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v28, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v26.l, v55.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v35.h
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v26.h, v54.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v27.l, v54.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v65.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v149, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v14.l, v102.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v14.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v85.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v83.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v55.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v50.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v149, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v15.l, v100.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v15.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v82.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v149, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v16.l, v98.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v16.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.h
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v39.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v149, v16
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v17.l, v97.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v17.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v70.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v68.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v38.h
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v38.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v149, v17
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v18.l, v87.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v18.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v66.h
+; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v149, v18
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v19.l, v85.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v19.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v149, v19
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v20.l, v83.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v20.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v55.l
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v54.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v34.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v149, v20
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v21.l, v81.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v21.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v33.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v149, v21
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v22.l, v71.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v22.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v32.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v28.h, v53.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v28.l, v53.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v33.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v33.h
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v28.h, v52.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v33.l, v29.h, v51.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v29.l, v50.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v149, v22
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v23.l, v70.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v23.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v51.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v149, v23
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v24.l, v67.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v24.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v149, v24
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v25.l, v66.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v25.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v149, v25
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v26.l, v64.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v26.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v149, v26
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v27.l, v54.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v27.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v149, v27
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v28.l, v53.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v28.l, v149.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v149, v28
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v29.l, v52.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v29.l, v149.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v30.l, v50.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v32.l, v49.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v149, v29
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v30.l, v51.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v30.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v149, v30
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v31.l, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v31.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v149, v31
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB38_2
; GFX11-TRUE16-NEXT: .LBB38_4: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v151.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v148.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v150.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v146.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v147.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v146.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, 0
; 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 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v149.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v151.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v150.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v134.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v31, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v148.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v149.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v147.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v134.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v148.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v145.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v31, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v31.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v145.h, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v131.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v31, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v130.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v130.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v144.h, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v145.l, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v135.h, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v135.l, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.l, v32.l, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v134.l, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v118.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v133.l, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v31, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v31, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v118.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v117.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v133.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v133.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v132.h, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v31, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v131.h, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v132.l, v6.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v132.l, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v117.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v114.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v31, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v113.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v129.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v129.h, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.l, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v112.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v128.l, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v119.h, v9.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v102.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v118.l, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v102.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v31, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v128.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v128.h, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v103.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v31, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v101.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v117.h, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v118.l, v9.h
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v117.h, v10.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v100.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v11.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v116.l, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v100.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v115.h, v11.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v97.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v114.h, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v31, v12
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v116.l, v10.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v116.h, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v99.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v31, v13
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v96.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v96.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v87.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v96.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v103.h, v13.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v85.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v103.l, v13.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v13.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v31, v14
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.h, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v113.l, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v86.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v31, v15
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v84.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v84.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v102.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v102.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v101.h, v14.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v83.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v15.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v101.l, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v84.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v99.h, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v31, v16
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v100.h, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.l, v14.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v82.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v99.l, v15.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v15.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v31, v17
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v80.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v80.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v98.h, v15.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v99.l, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v98.l, v16.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v71.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v97.l, v16.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v71.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v16.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v16.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v87.l, v17.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v31, v18
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v97.l, v16.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v97.h, v16.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v17.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v31, v19
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v68.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v67.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v87.l, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v87.h, v17.h
; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v86.h, v18.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v67.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v85.l, v18.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v67.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v18.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v84.h, v19.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.h, v19.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v31, v20
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v85.l, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v85.h, v18.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v19.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v65.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v31, v21
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v55.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v50.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v83.l, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v83.h, v19.h
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v82.h, v20.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v51.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v21.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v82.l, v20.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v52.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v20.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v80.h, v21.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v22.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v80.l, v21.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v21.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v31, v22
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v81.h, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v82.l, v20.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v31, v23
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v48.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.l, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v71.h, v21.h
; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v70.h, v22.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v39.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.l, v22.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v22.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v22.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v68.h, v23.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v68.l, v23.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v23.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v31, v24
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v70.l, v22.l
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v70.h, v22.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v31, v25
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v38.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v67.h, v23.l
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v68.l, v23.h
; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v37.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v66.h, v24.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v64.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v26, v31
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v65.l, v25.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v25.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v26
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v25.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.l, v24.l
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v66.h, v24.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v27
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v25.h
+; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v36.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v64.l, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v64.h, v25.h
; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l
; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v27, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v55.l, v26.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v55.h, v26.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v35.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v28
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v27.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v28.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v53.h, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v31, v28
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v27.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v26.l
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v55.l, v26.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v31, v29
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v34.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v53.h, v27.l
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v54.l, v27.h
; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v33.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v28.h
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v31, v30
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v34.h, 0x300, v29.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v28.l
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v53.l, v28.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v53.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v30.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v31, v34
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v29.h
+; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v28.l
-; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v32.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v51.h, v29.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v52.l, v29.h
; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.l, v30.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v33
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v50.l, v30.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.l, 0x300, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v30.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v49.h, v30.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v32.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v31, v33
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v32.h
+; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v50.h, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v51.l, v30.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v31, v33
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.h, 0x300, v32.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v32
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -77938,64 +77682,64 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2)
@@ -78023,50 +77767,50 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
@@ -78135,50 +77879,50 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT: .LBB56_4: ; %end
@@ -78186,307 +77930,266 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v1.h, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff, v66
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v55, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v150.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v66, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v55, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v4.l, v4.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v55, v39
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v149.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l
+; 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_or_b32_e32 v4, v39, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v148.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v147.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v55, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v55, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v135.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v8.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v133.l
; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l
; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v132.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v53, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v131.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v53, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v119.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v116.l
; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v129.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v114.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v113.l
; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v53
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v51.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v53
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v118.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v51, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v102.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v51, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v116.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v115.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v99.l
; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v50, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v114.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v50, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v113.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v96.l
; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l
; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v49, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v112.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l
+; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v85.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l
; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v50
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v18.l, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v50
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v49, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v19.l, v19.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v20.l
-; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v101.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v48, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v49
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v82.l
; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v81.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v71.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v48, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v22.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v49
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v21.l, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v98.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l
; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v22.l, v22.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v97.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v69.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v87.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v24.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v25.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v84.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v27.l, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v81.l
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v28.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v80.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v71.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v70.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v69.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v30.l, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v39
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v31.l, v31.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l
-; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v68.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v32.l, v32.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v39
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32
; GFX11-TRUE16-NEXT: s_clause 0x5
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
@@ -87060,63 +86763,63 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:384
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:372
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:368
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:364
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:360
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:356
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:348
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:344
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:360
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:352
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:344
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:324
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:316
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:264
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:288
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:280
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:272
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:264
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260
; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:256
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:256
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:240
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:240
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v103, off, s32 offset:388
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:232
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:228
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:224
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:220
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:216
+; GFX11-TRUE16-NEXT: scratch_load_b32 v114, off, s32 offset:388
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:16
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:16
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:80
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:32
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:40
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:48
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:56
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:64
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:72
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:80
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:96
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:96
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:112
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:112
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v161, off, s32 offset:136
@@ -87130,144 +86833,143 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:212
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:204
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:196
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:188
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:180
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:172
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:164
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:140
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:132
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:116
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:100
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:84
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:52
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:44
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:20
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v116, off, s32 offset:4
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.h, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v12.l
; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.l, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v29.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v29.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54)
-; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v103
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(53)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v80.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v81.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(52)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v82.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v82.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v82.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v83.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(50)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v83.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v83.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(49)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v84.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v85.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(48)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.h, 8, v85.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v85.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v86.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(46)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v87.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v87.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(45)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v96.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v97.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(44)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.h, 8, v97.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v97.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v98.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(42)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v99.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v99.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(41)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v99.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(40)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v101.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v101.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v102.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(38)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v160.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(37)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v160.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(36)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v161.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.h, 8, v161.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v161.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v161.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(34)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v162.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v162.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v162.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v162.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(32)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v163.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v163.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(30)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v164.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v164.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v164.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(28)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v165.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v165.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v70.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v70.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v68.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v65.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v165.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v71.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v71.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v70.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v67.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v66.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v64.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.l, 8, v55.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v54.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v54.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v53.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v52.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v51.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v49.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v54.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v53.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v52.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v51.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v50.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v31.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
@@ -87281,746 +86983,660 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB58_3: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v151.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v148.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v150.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v146.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v151.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v150.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v147.h
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v0.h, v149.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v149.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v144.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v133.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v148.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v144.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v145.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v1.h, v147.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v146.l
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v0.l, v151.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v151.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v0.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v150.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v149, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v1.h, v150.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v132.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v131.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v148.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v130.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v149, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v2.l, v148.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v149.h
; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v131.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v2.l, v146.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v131.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v135.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v3.l, v145.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v128.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v118.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v134.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v145.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v130.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v149, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v3.l, v147.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v147.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v118.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v149, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v4.l, v144.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v4.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v133.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v117.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v116.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v4.l, v135.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v134.l
-; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v5.l, v133.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v132.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v114.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v6.l, v132.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v130.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v7.l, v130.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v129.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v102.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v8.l, v128.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v119.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v9.l, v118.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v102.h
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v117.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v97.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v10.l, v116.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v100.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v115.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v11.l, v114.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v11.h, v113.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v85.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v12.l, v113.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v103.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v13.l, v103.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v101.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v81.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v14.l, v101.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v14.h, v99.h
-; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v71.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v15.l, v99.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v15.h, v98.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v16.l, v97.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v71.h
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v16.h, v96.h
-; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v67.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v17.l, v87.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v149, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v5.l, v135.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v5.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v132.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v129.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v113.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v112.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v149, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v6.l, v133.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v6.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v128.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v103.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v101.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v149, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v7.l, v131.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v7.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v100.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v99.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v149, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v8.l, v129.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v116.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v114.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v96.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v149, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v9.l, v128.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v9.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v113.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v86.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v149, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v10.l, v117.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v10.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v102.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v84.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v82.h
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v81.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v149, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v11.l, v116.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v11.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.l
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v99.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v149, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v12.l, v114.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v12.l, v149.h
; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v17.h, v86.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v64.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v18.l, v85.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v19.l
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v67.h
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v18.h, v84.h
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v51.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v19.l, v83.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v97.h
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v69.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v68.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v149, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v13.l, v112.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v13.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v87.h
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v67.l
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v19.h, v82.h
-; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v48.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v20.l, v82.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v20.h, v80.h
-; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v39.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v21.l, v80.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v21.h, v70.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v22.l, v70.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v22.h, v68.h
-; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v38.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v23.l, v68.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v38.h
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v23.h, v66.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v36.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v24.l, v66.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v37.h
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v24.h, v65.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v36.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v27, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v25.l, v64.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v25.h, v55.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v28, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v26.l, v55.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v35.h
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v26.h, v54.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v27.l, v54.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v65.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v149, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v14.l, v102.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v14.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v85.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v83.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v55.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v50.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v149, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v15.l, v100.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v15.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v82.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v149, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v16.l, v98.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v16.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.h
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v39.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v149, v16
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v17.l, v97.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v17.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v70.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v68.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v38.h
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v38.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v149, v17
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v18.l, v87.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v18.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v66.h
+; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v149, v18
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v19.l, v85.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v19.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v149, v19
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v20.l, v83.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v20.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v55.l
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v54.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v34.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v149, v20
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v21.l, v81.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v21.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v33.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v149, v21
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v22.l, v71.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v22.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v32.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v28.h, v53.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v28.l, v53.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v33.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v33.h
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v28.h, v52.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v33.l, v29.h, v51.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v29.l, v50.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v149, v22
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v23.l, v70.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v23.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v51.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v149, v23
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v24.l, v67.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v24.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v149, v24
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v25.l, v66.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v25.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v149, v25
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v26.l, v64.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v26.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v149, v26
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v27.l, v54.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v27.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v149, v27
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v28.l, v53.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v28.l, v149.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v149, v28
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v29.l, v52.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v29.l, v149.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v30.l, v50.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v32.l, v49.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v149, v29
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v30.l, v51.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v30.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v149, v30
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v31.l, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v31.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v149, v31
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB58_2
; GFX11-TRUE16-NEXT: .LBB58_4: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v151.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v148.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v150.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v146.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v147.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v146.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, 0
; 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 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v149.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v151.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v150.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v134.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v31, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v148.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v149.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v147.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v134.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v148.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v145.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v31, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v31.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v145.h, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v131.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v31, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v130.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v130.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v144.h, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v145.l, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v135.h, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v135.l, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.l, v32.l, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v134.l, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v118.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v133.l, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v31, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v31, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v118.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v117.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v133.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v133.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v132.h, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v31, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v131.h, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v132.l, v6.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v132.l, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v117.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v114.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v31, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v113.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v129.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v129.h, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.l, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v112.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v128.l, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v119.h, v9.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v102.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v118.l, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v102.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v31, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v128.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v128.h, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v103.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v31, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v101.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v117.h, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v118.l, v9.h
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v117.h, v10.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v100.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v11.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v116.l, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v100.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v115.h, v11.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v97.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v114.h, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v31, v12
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v116.l, v10.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v116.h, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v99.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v31, v13
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v96.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v96.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v87.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v96.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v103.h, v13.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v85.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v103.l, v13.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v13.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v31, v14
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.h, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v113.l, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v86.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v31, v15
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v84.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v84.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v102.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v102.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v101.h, v14.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v83.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v15.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v101.l, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v84.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v99.h, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v31, v16
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v100.h, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.l, v14.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v82.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v99.l, v15.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v15.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v31, v17
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v80.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v80.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v98.h, v15.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v99.l, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v98.l, v16.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v71.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v97.l, v16.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v71.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v16.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v16.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v87.l, v17.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v31, v18
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v97.l, v16.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v97.h, v16.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v17.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v31, v19
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v68.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v67.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v87.l, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v87.h, v17.h
; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v86.h, v18.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v67.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v85.l, v18.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v67.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v18.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v84.h, v19.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.h, v19.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v31, v20
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v85.l, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v85.h, v18.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v19.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v65.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v31, v21
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v55.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v50.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v83.l, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v83.h, v19.h
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v82.h, v20.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v51.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v21.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v82.l, v20.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v52.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v20.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v80.h, v21.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v22.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v80.l, v21.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v21.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v31, v22
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v81.h, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v82.l, v20.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v31, v23
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v48.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.l, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v71.h, v21.h
; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v70.h, v22.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v39.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.l, v22.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v22.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v22.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v68.h, v23.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v68.l, v23.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v23.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v31, v24
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v70.l, v22.l
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v70.h, v22.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v31, v25
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v38.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v67.h, v23.l
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v68.l, v23.h
; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v37.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v66.h, v24.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v64.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v26, v31
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v65.l, v25.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v25.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v26
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v25.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.l, v24.l
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v66.h, v24.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v27
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v25.h
+; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v36.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v64.l, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v64.h, v25.h
; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l
; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v27, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v55.l, v26.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v55.h, v26.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v35.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v28
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v27.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v28.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v53.h, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v31, v28
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v27.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v26.l
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v55.l, v26.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v31, v29
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v34.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v53.h, v27.l
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v54.l, v27.h
; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v33.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v28.h
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v31, v30
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v34.h, 0x300, v29.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v28.l
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v53.l, v28.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v53.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v30.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v31, v34
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v29.h
+; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v28.l
-; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v32.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v51.h, v29.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v52.l, v29.h
; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.l, v30.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v33
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v50.l, v30.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.l, 0x300, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v30.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v49.h, v30.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v32.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v31, v33
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v32.h
+; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v50.h, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v51.l, v30.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v31, v33
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.h, 0x300, v32.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v32
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -111800,64 +111416,64 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2)
@@ -111885,50 +111501,50 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
@@ -111972,50 +111588,50 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT: .LBB72_4: ; %end
@@ -112023,307 +111639,266 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v1.h, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v66, 0xffff, v66
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v55, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v65.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v150.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v66, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v55, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v4.l, v4.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v55, v39
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v149.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l
+; 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_or_b32_e32 v4, v39, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v148.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v147.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v65
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v55, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v55, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v54.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v55, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v135.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v8.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v133.l
; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l
; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v132.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v53, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v53, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v131.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v53, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v119.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v116.l
; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v129.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v114.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v113.l
; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v53
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v51.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v53
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v118.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v51, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v102.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v51, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v116.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v115.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v99.l
; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v50, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v50, 0xffff, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v114.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v50, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v113.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v96.l
; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l
; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v49, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v112.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l
+; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v85.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l
; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v50
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v18.l, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v50
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v49, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v19.l, v19.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v20.l
-; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v101.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v48, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v49
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v82.l
; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v81.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v71.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v48, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v22.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v49
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v21.l, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v98.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l
; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v22.l, v22.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v97.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v69.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v87.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v24.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v25.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v84.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v27.l, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v81.l
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v28.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v80.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v71.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v70.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v69.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v30.l, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v39
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v31.l, v31.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l
-; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v68.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v39
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v39.h, v32.l, v32.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v39
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32
; GFX11-TRUE16-NEXT: s_clause 0x5
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
@@ -121839,63 +121414,63 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:384
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:372
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:368
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:364
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:360
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:356
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:348
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:344
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:360
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:352
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:344
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:324
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:316
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:292
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:288
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:284
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:280
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:276
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:272
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:268
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:264
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:288
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:280
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:272
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:264
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260
; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:256
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:256
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:240
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:240
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:232
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:228
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:224
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:220
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:216
-; GFX11-TRUE16-NEXT: scratch_load_b32 v103, off, s32 offset:388
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:232
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:228
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:224
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:220
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:216
+; GFX11-TRUE16-NEXT: scratch_load_b32 v114, off, s32 offset:388
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:16
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:16
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:80
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:32
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:40
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:48
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:56
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:64
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:72
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:80
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:96
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:96
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:112
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:112
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v161, off, s32 offset:136
@@ -121909,144 +121484,143 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:212
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:204
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:196
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:188
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:180
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:172
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:164
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:212
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:204
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:196
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:188
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:180
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:172
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:164
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:140
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:132
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:116
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:100
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v114, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:84
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:52
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:44
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:20
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v116, off, s32 offset:4
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.h, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v12.l
; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v149.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.l, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.h, 8, v29.l
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v29.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54)
-; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v103
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(53)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v80.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v81.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(52)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v82.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v82.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(51)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v82.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v83.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(50)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v83.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v83.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(49)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v84.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v85.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(48)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.h, 8, v85.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v85.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(47)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v86.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(46)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v87.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v87.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(45)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v96.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v97.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(44)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.h, 8, v97.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v97.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(43)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v98.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(42)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v99.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v99.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(41)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.l, 8, v99.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(40)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v103.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v101.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(39)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v101.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v102.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(38)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v160.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(37)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v160.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(36)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.h, 8, v161.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.h, 8, v161.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(35)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v161.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v161.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(34)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v162.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v162.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v162.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v162.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(32)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v163.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v163.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(30)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v86.h, 8, v164.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v164.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v164.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(28)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v165.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v165.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v70.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v70.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v68.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v65.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v165.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v71.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v71.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v70.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v67.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v66.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v64.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v65.l, 8, v55.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v54.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v54.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v53.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v52.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v51.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v49.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v54.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v53.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v52.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v51.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v50.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v31.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
@@ -122060,746 +121634,660 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB74_3: ; %cmp.false
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v151.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v148.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v150.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v146.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v151.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v150.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v147.h
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v0.h, v149.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v149.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v144.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v133.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v148.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v144.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v145.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v1.h, v147.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v146.l
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v0.l, v151.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v151.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v0.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v150.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v149, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v1.h, v150.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v132.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v131.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v148.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v130.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v149, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v2.l, v148.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v149.h
; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v131.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v2.l, v146.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v131.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v135.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v3.l, v145.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v128.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v118.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v134.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v145.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v130.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v149, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v3.l, v147.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v147.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v118.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v149, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v4.l, v144.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v4.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v133.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v117.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v116.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v4.l, v135.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v134.l
-; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v5.l, v133.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v132.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v114.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v6.l, v132.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v130.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v7.l, v130.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v129.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v102.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v8.l, v128.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v119.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v9.l, v118.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v102.h
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v117.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v97.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v10.l, v116.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v100.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v115.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v11.l, v114.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v11.h, v113.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v85.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v12.l, v113.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v103.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v13.l, v103.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v101.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v81.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v14.l, v101.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v14.h, v99.h
-; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v71.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v15.l, v99.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v15.h, v98.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v16.l, v97.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v71.h
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v16.h, v96.h
-; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v67.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v17.l, v87.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v149, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v5.l, v135.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v5.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v132.l
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v129.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v113.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v112.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v149, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v6.l, v133.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v6.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v128.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v103.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v101.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v149, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v7.l, v131.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v7.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v100.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v99.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v149, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v8.l, v129.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v116.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v114.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v96.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v149, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v9.l, v128.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v9.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v113.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v86.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v149, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v10.l, v117.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v10.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v102.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v84.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v82.h
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v81.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v149, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v11.l, v116.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v11.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.l
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v99.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v149, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v12.l, v114.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v12.l, v149.h
; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v17.h, v86.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v64.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v18.l, v85.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v19.l
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v67.h
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v18.h, v84.h
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v51.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v19.l, v83.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v97.h
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v69.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v68.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v149, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v13.l, v112.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v13.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v87.h
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v67.l
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v19.h, v82.h
-; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v48.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v20.l, v82.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v20.h, v80.h
-; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v39.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v21.l, v80.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v21.h, v70.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v22.l, v70.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v22.h, v68.h
-; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v38.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v23.l, v68.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v38.h
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v23.h, v66.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v36.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v24.l, v66.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v37.h
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v24.h, v65.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v36.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v27, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v28
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v25.l, v64.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v25.h, v55.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v28, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v26.l, v55.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v35.h
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v26.h, v54.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v27.l, v54.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v65.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v149, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v14.l, v102.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v14.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v85.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v83.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v55.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v50.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v149, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v15.l, v100.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v15.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v82.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v149, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v16.l, v98.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v16.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.h
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v39.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v149, v16
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v17.l, v97.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v17.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v70.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v68.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v38.h
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v38.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v149, v17
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v18.l, v87.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v18.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v66.h
+; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v149, v18
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v19.l, v85.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v19.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v149, v19
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v20.l, v83.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v20.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v55.l
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v54.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v34.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v149, v20
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v21.l, v81.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v21.l, v149.h
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v33.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v149, v21
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v22.l, v71.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v22.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v32.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v28.h, v53.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v34.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v28.l, v53.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v33.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v33.h
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v28.h, v52.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v33.l, v29.h, v51.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v34
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v29.l, v50.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v149, v22
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v23.l, v70.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v23.l, v149.h
+; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v51.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v149, v23
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v24.l, v67.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v24.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v149, v24
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v25.l, v66.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v25.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v149, v25
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v26.l, v64.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v26.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v149, v26
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v27.l, v54.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v27.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v149, v27
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v28.l, v53.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v28.l, v149.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v149, v28
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v29.l, v52.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v29.l, v149.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v30.l, v50.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v32.l, v49.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v149, v29
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v30.l, v51.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v30.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v149, v30
+; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v31.l, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v31.l, v149.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v149, v31
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB74_2
; GFX11-TRUE16-NEXT: .LBB74_4: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v151.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v148.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v150.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v146.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v147.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v146.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, 0
; 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 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v144.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v149.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v151.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v150.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v134.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v31, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v148.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v149.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v147.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v134.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v148.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v145.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v31, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v31.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v145.h, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v131.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v31, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v130.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v130.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v144.h, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v145.l, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v135.h, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v135.l, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.l, v32.l, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v134.l, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v118.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v133.l, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v31, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v31, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v118.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v117.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v133.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v133.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v132.h, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v31, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v131.h, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v132.l, v6.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v132.l, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v117.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v114.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v31, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v113.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v129.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v129.h, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v129.l, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v112.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v128.l, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v119.h, v9.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v102.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v118.l, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v102.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v31, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v128.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v128.h, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v103.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v31, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v101.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v117.h, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v118.l, v9.h
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v117.h, v10.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v100.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v11.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v116.l, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v100.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v115.h, v11.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v97.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v114.h, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v13
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v31, v12
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v116.l, v10.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v116.h, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v99.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v31, v13
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v96.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v96.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v87.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v14, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v96.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v103.h, v13.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v85.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v103.l, v13.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v13.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v31, v14
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.h, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v113.l, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v86.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v31, v15
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v84.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v84.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v102.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v102.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v101.h, v14.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v83.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v15.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v101.l, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v84.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v99.h, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v31, v16
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v100.h, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.l, v14.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v82.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v16.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v99.l, v15.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v15.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v31, v17
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v80.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v80.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v98.h, v15.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v99.l, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v98.l, v16.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v71.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v18, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v97.l, v16.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v71.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v16.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v16.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v19, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v87.l, v17.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v31, v18
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v97.l, v16.l
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v97.h, v16.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v17.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v31, v19
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v68.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v67.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v87.l, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v87.h, v17.h
; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v86.h, v18.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v67.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v20, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v85.l, v18.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v67.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v18.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v84.h, v19.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v21, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v83.h, v19.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v31, v20
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v85.l, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v85.h, v18.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v19.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v65.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v31, v21
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v55.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v50.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v83.l, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v83.h, v19.h
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v82.h, v20.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v51.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v21.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v82.l, v20.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v52.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v20.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v80.h, v21.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v23, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v22.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v80.l, v21.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v21.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v31, v22
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v81.h, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v82.l, v20.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v31, v23
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v48.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.l, v21.l
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v71.h, v21.h
; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v70.h, v22.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v39.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v24, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.l, v22.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v22.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v22.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v68.h, v23.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v25, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v68.l, v23.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v23.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v31, v24
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v70.l, v22.l
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v70.h, v22.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v31, v25
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v38.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v67.h, v23.l
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v68.l, v23.h
; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v37.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v26, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v66.h, v24.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v26
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v64.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v24.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v26, v31
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v65.l, v25.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v27, 0xffff, v27
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v25.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v26
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v25.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.l, v24.l
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v66.h, v24.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v27
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v25.h
+; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v36.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v64.l, v25.l
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v64.h, v25.h
; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l
; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v27, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v55.l, v26.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v55.h, v26.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v35.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v28
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v29, v31
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v27.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v28.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h
-; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v29, 0xffff, v29
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v53.h, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v31, v28
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v27.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v26.l
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v55.l, v26.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v31, v29
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v34.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v53.h, v27.l
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v54.l, v27.h
; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v29, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v33.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v28.h
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v31, v30
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v34.h, 0x300, v29.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v28.l
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v53.l, v28.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v53.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v30.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v31, v34
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v29.h
+; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v28.l
-; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v32.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v51.h, v29.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v52.l, v29.h
; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.l, v30.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v33
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v50.l, v30.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.l, 0x300, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v31
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v30.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v33
-; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v49.h, v30.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v31
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v32.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v31
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v31, v33
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v32.h
+; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v50.h, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v51.l, v30.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v31, v33
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.h, 0x300, v32.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v32
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -160089,159 +159577,162 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:156
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:136
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:32
-; GFX11-TRUE16-NEXT: s_clause 0x4
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:12
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:168
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:164
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:160
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:156
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:152
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:148
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:144
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:140
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:136
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:128
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:124
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:120
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:112
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:108
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:104
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:96
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:92
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:88
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:84
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:80
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:76
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:72
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:64
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:56
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:52
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:48
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:44
+; GFX11-TRUE16-NEXT: s_clause 0x7
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:40
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:24
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v109, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v110, s32 offset:16
+; GFX11-TRUE16-NEXT: scratch_store_b32 off, v111, s32 offset:12
; GFX11-TRUE16-NEXT: s_clause 0x2
; GFX11-TRUE16-NEXT: scratch_load_b32 v33, off, s32 offset:8
; GFX11-TRUE16-NEXT: scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr111_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr106_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr108_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr104_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr107_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr164_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr106_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr94_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr90_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr180_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr165_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr95_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr93_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr166_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr88_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr75_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr47_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr76_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr179_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr72_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr43_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr74_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr177_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr63_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr178_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr59_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr60_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr73_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr58_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr57_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr44_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr56_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr41_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr42_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr89_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr43_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr61_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr183_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr57_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr167_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr104_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr176_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr166_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr77_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr95_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr93_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr47_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr44_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr92_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr79_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr40_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr74_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr62_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr63_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr59_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr182_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr62_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr180_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr108_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr176_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr60_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr46_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr165_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr89_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr110_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr40_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr107_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr109_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr181_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr182_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr177_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr94_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr90_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr79_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr77_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr75_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr72_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr61_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr58_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr56_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr46_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr42_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr183_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr181_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr179_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr167_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr164_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
@@ -160250,143 +159741,142 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB90_2
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[7:8]
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[5:6]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[7:8]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[11:12]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[3:4]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[5:6]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v42, 24, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v56, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v59, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v76, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v75, 24, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v88, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v91, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 24, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v94, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v106, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v180, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v40, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v47, 8, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v57, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v60, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 8, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v74, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v76, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v78, 8, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v88, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v93, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v95, 8, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v104, 8, v3
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v105, 24, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v107, 8, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v108, 8, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 24, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v27
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 24, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v177, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 24, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v40, 8, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 8, v21
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v46, 24, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v60, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 24, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v74, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v79, 8, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18]
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.h, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.h, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v164.h, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.h, v4.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v180.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v165.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v161.h, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v47.h, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v179.h, v8.l
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v106, 8, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v111, 8, v1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 8, v27
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 24, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v179, 8, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v42, 8, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v46, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v56, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v61, 8, v21
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 24, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v75, 8, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v77, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v79, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 8, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v94, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[15:16]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[1:2]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[23:24]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[21:22]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18]
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.h, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.h, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v166.h, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.h, v6.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v43.h, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v177.h, v8.l
; GFX11-TRUE16-NEXT: v_mov_b16_e64 v178.h, v8.h
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v73.h, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v44.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v41.h, v10.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v89.h, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v61.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v57.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v104.h, v13.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v78.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v77.h, v14.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v95.h, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v93.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v92.h, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.h, v17.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.h, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v18.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v84.h, v19.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.h, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v20.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.h, v21.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.h, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.h, v22.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.h, v23.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v87.h, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.h, v24.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v101.h, v25.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v98.h, v26.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.h, v26.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v112.h, v27.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v100.h, v28.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v99.h, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v113.h, v29.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v103.h, v30.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v102.h, v30.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.h, v31.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v115.h, v32.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v114.h, v32.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v41.h, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v44.h, v10.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v92.h, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v59.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v62.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v108.h, v13.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v91.h, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v89.h, v14.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v110.h, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v107.h, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v109.h, v16.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.h, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.h, v19.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.h, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v84.h, v20.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.h, v21.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.h, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v87.h, v22.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v99.h, v23.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.h, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v98.h, v24.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v102.h, v25.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v100.h, v26.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v101.h, v26.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v113.h, v27.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v103.h, v28.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v112.h, v28.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.h, v29.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v114.h, v30.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v115.h, v30.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v31.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.h, v32.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v32.h
; GFX11-TRUE16-NEXT: .LBB90_2: ; %Flow
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB90_4
; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff0000, v18
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v18, 16, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v20
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v18, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v18
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add3_u32 v37, v37, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v70, v37, v39 :: v_dual_add_f32 v33, 0x40c00000, v33
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v80, v37, v39 :: v_dual_add_f32 v33, 0x40c00000, v33
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_bfe_u32 v36, v33, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT: v_add3_u32 v36, v36, v33, 0x7fff
; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v17
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v70.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v80.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v55, v36, v38, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v81, v36, v38, vcc_lo
; GFX11-TRUE16-NEXT: v_dual_add_f32 v34, 0x40c00000, v34 :: v_dual_lshlrev_b32 v17, 16, v17
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
@@ -160399,498 +159889,500 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_add3_u32 v18, v48, v34, 0x7fff
; GFX11-TRUE16-NEXT: v_add3_u32 v37, v50, v17, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v71, v37, v51 :: v_dual_lshlrev_b32 v20, 16, v20
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v20, 0x40c00000, v20
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v82, v37, v51 :: v_dual_and_b32 v35, 0xffff0000, v20
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v20, 16, v20
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-TRUE16-NEXT: v_and_b32_e32 v51, 0xffff0000, v11
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v71.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v82.h
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-TRUE16-NEXT: v_bfe_u32 v36, v35, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v20
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v18, v49, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfi_b32 v18, 0xffff, v33, v55
+; GFX11-TRUE16-NEXT: v_bfi_b32 v18, 0xffff, v33, v81
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v35
; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v20, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v20
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v35
; GFX11-TRUE16-NEXT: v_bfi_b32 v17, 0xffff, v34, v17
; GFX11-TRUE16-NEXT: v_add3_u32 v34, v36, v35, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v19
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19
; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 24, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v74, 8, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_add_f32 v19, 0x40c00000, v19
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v81, v33, v37, vcc_lo
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v19
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v79, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 8, v18
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v83, v33, v37, vcc_lo
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_lshlrev_b32 v19, 16, v19
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v22
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22
-; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v19, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_lshlrev_b32 v22, 16, v22
; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v36, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v80, v34, v38, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v84, v34, v38, vcc_lo
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v83.h
+; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v19, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v22, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v19, 0x7fff
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v36, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v81.h
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v79, 8, v17
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v84, v33, v37, vcc_lo
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v94, 8, v17
+; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v19, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v22, 0x40c00000, v22 :: v_dual_cndmask_b32 v85, v33, v37
; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v22, 16, 1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v84.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v85.h
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v20, v38, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v35, 16, 1
; GFX11-TRUE16-NEXT: v_add3_u32 v20, v33, v22, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v33, 0x400000, v22
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v86, v20, v33 :: v_dual_add_f32 v35, 0x40c00000, v35
+; GFX11-TRUE16-NEXT: v_bfi_b32 v20, 0xffff, v34, v84
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v86.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v35, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v35
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v35, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v83, v20, v33, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT: v_bfi_b32 v20, 0xffff, v34, v80
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v82, v19, v39, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfi_b32 v19, 0xffff, v37, v36
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v24
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v24, 16, v24
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v83.h
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v46, 24, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v60, 8, v20
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
-; GFX11-TRUE16-NEXT: v_bfi_b32 v22, 0xffff, v22, v82
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 8, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v36
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 24, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v75, 8, v20
+; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v35, 0x7fff
; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v21
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 24, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v40, 8, v22
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v33, 0x40c00000, v38
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v87, v19, v39, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfi_b32 v19, 0xffff, v37, v36
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v24
; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v33, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v33, 0x40c00000, v38 :: v_dual_lshlrev_b32 v24, 16, v24
+; GFX11-TRUE16-NEXT: v_bfi_b32 v22, 0xffff, v22, v87
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v21, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v33, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v21
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v24, 0x40c00000, v24
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33
-; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v33, 0x7fff
; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v21, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v86, v34, v37 :: v_dual_and_b32 v37, 0xffff0000, v23
+; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v33, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v36
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v56, 24, v22
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v96, v34, v37 :: v_dual_and_b32 v37, 0xffff0000, v23
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v24, 16, 1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v23, 16, v23
+; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v24, 16, 1
; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v36, 16, 1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v22
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v21, v35, v38, vcc_lo
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v24, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v24
+; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v24, 0x7fff
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v23, 0x40c00000, v23
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v36, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v86.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v87, v34, v38, vcc_lo
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v26
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v23, 0x40c00000, v23 :: v_dual_lshlrev_b32 v26, 16, v26
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v96.h
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v77, 8, v19
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v97, v34, v38, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT: v_bfi_b32 v21, 0xffff, v35, v21
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v26
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v23, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v85, v33, v39, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v98, v33, v39, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v37, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, 0x400000, v23
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 8, v21
; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v23, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v97, v34, v36, vcc_lo
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v37, 0x40c00000, v37 :: v_dual_add_f32 v34, 0x40c00000, v38
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v97.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v37, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v37
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34
; GFX11-TRUE16-NEXT: v_add3_u32 v24, v24, v37, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v34, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v26, 0x40c00000, v26
+; GFX11-TRUE16-NEXT: v_bfi_b32 v21, 0xffff, v35, v21
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v99, v34, v36, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v34, 0x40c00000, v38
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v61, 8, v21
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v99.h
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v23, v24, v39, vcc_lo
-; GFX11-TRUE16-NEXT: v_add3_u32 v35, v37, v34, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v26
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v34, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34
; GFX11-TRUE16-NEXT: v_bfi_b32 v23, 0xffff, v36, v23
; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v25
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v87.h
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v23
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v97.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25
+; GFX11-TRUE16-NEXT: v_add3_u32 v35, v37, v34, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v26
; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT: v_bfi_b32 v24, 0xffff, v33, v85
+; GFX11-TRUE16-NEXT: v_bfi_b32 v24, 0xffff, v33, v98
; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v26, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 24, v24
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v25, 0x40c00000, v25
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v46, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 24, v24
; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v26, 0x7fff
; GFX11-TRUE16-NEXT: v_bfe_u32 v26, v36, 16, 1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v177, 8, v24
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v98, v33, v37, vcc_lo
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v42, 8, v24
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v100, v33, v37, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v25, 16, v25
-; GFX11-TRUE16-NEXT: v_add3_u32 v26, v26, v36, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v96, v35, v38 :: v_dual_add_f32 v25, 0x40c00000, v25
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v28
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v28, 16, v28
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v98.h
; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v25, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v25
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v28, 0x40c00000, v28
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_add3_u32 v26, v26, v36, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v101, v35, v38, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v25, 0x7fff
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v28
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v100.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v102, v33, v37, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v28, 16, v28
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v102.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v35, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v35
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v101, v33, v37, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v26, v38, vcc_lo
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v27
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v27
; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v28, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-TRUE16-NEXT: v_add3_u32 v25, v25, v35, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v26, v38, vcc_lo
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v35
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
; GFX11-TRUE16-NEXT: v_add3_u32 v26, v33, v28, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v33, 0x400000, v28
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v27
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v27
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v101.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v100, v26, v33, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v103, v26, v33, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v27, 0x40c00000, v27
-; GFX11-TRUE16-NEXT: v_bfi_b32 v26, 0xffff, v34, v96
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v100.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v99, v25, v39, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfi_b32 v25, 0xffff, v37, v36
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v30
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v33, 0x40c00000, v38
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; GFX11-TRUE16-NEXT: v_bfi_b32 v26, 0xffff, v34, v101
; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v27, 16, 1
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v33, 0x40c00000, v38
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v103.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v112, v25, v39, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfi_b32 v25, 0xffff, v37, v36
+; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v27, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v27
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v33, 16, 1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27
-; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v27, 0x7fff
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v30
+; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v33, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
+; GFX11-TRUE16-NEXT: v_bfi_b32 v28, 0xffff, v28, v112
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v113, v34, v37, vcc_lo
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff0000, v29
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v30
; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v33, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v36
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v112, v34, v37, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v36, 16, 1
-; GFX11-TRUE16-NEXT: v_bfi_b32 v28, 0xffff, v28, v99
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v26
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_add_f32 v37, 0x40c00000, v37
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v30, 0x40c00000, v30 :: v_dual_lshlrev_b32 v29, 16, v29
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v27, v35, v38, vcc_lo
-; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v36, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff0000, v29
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v29, 16, v29
+; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v36, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v36
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v30, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v30
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v112.h
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v28
-; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v30, 0x7fff
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v25
-; GFX11-TRUE16-NEXT: v_bfi_b32 v27, 0xffff, v35, v27
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v103, v34, v38 :: v_dual_and_b32 v38, 0xffff0000, v32
+; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v36, 0x7fff
; GFX11-TRUE16-NEXT: v_add_f32_e32 v29, 0x40c00000, v29
+; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v30, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v30, v37, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v113.h
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v28
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v114, v34, v38, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v32
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v27
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v29, 16, 1
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v102, v33, v39 :: v_dual_add_f32 v37, 0x40c00000, v37
; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, 0x400000, v29
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v115, v33, v39, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
+; GFX11-TRUE16-NEXT: v_add3_u32 v30, v30, v37, 0x7fff
; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v29, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v103.h
-; GFX11-TRUE16-NEXT: v_bfe_u32 v30, v37, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v37
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v113, v34, v36, vcc_lo
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v114.h
+; GFX11-TRUE16-NEXT: v_bfi_b32 v27, 0xffff, v35, v27
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 24, v26
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v116, v34, v36, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT: v_add3_u32 v30, v30, v37, 0x7fff
; GFX11-TRUE16-NEXT: v_add_f32_e32 v34, 0x40c00000, v38
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v113.h
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 8, v27
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v179, 8, v26
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v116.h
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v29, v30, v39, vcc_lo
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v34, 16, 1
-; GFX11-TRUE16-NEXT: v_bfi_b32 v30, 0xffff, v33, v102
-; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v32, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
+; GFX11-TRUE16-NEXT: v_bfi_b32 v30, 0xffff, v33, v115
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 8, v25
; GFX11-TRUE16-NEXT: v_bfi_b32 v29, 0xffff, v36, v29
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v31
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_lshlrev_b32 v31, 16, v31
; GFX11-TRUE16-NEXT: v_add3_u32 v35, v37, v34, 0x7fff
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v30
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
+; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v32, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v32
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v29
; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v32, 0x7fff
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v31
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v30
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v115, v33, v37, vcc_lo
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_lshlrev_b32 v31, 16, v31
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v117, v33, v37, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v29
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v115.h
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v31, 0x40c00000, v31
-; GFX11-TRUE16-NEXT: v_bfe_u32 v32, v36, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v114, v35, v38, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v31, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v31
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-TRUE16-NEXT: v_add3_u32 v32, v32, v36, 0x7fff
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v117.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v31, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v116, v33, v37 :: v_dual_and_b32 v35, 0xffff0000, v2
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v2, 16, v2
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v118, v35, v38, vcc_lo
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v2
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v35, 0x40c00000, v35
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v32, v36, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v119, v33, v37, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v116.h
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_bfe_u32 v31, v35, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v35
+; GFX11-TRUE16-NEXT: v_add3_u32 v32, v32, v36, 0x7fff
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v119.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add3_u32 v31, v31, v35, 0x7fff
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v32, v38, vcc_lo
; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v1, 16, v1
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v2, 16, 1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT: v_add3_u32 v31, v31, v35, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v35
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT: v_add3_u32 v32, v33, v2, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v33, 0x400000, v2
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v133, v32, v33, vcc_lo
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v33, 0x40c00000, v38
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v128, v32, v33, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT: v_bfi_b32 v32, 0xffff, v34, v114
+; GFX11-TRUE16-NEXT: v_bfi_b32 v32, 0xffff, v34, v118
; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v1, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v132, v31, v39, vcc_lo
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v33, 0x40c00000, v38
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v128.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v129, v31, v39, vcc_lo
; GFX11-TRUE16-NEXT: v_bfi_b32 v31, 0xffff, v37, v36
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v33, 16, 1
; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v1, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v33, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33
-; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v33, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v133.h
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 24, v32
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v146, v34, v37 :: v_dual_and_b32 v37, 0xffff0000, v3
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v131, v34, v37, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v132
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v37, 0x40c00000, v37
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
+; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v33, 0x7fff
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff0000, v3
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_lshlrev_b32 v3, 16, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v4, 16, 1
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v1, v35, v38 :: v_dual_add_f32 v36, 0x40c00000, v36
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v35, v38, vcc_lo
; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v4
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v4, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v36, 16, 1
+; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v4, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v36
-; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v37, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v148, v34, v38, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v35.l, v131.h
+; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v129
; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v36, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v133, v34, v38, vcc_lo
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v6
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v35, v1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v105, 24, v2
+; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v135, v33, v39, vcc_lo
; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, 0x400000, v3
-; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v37, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v3, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v6
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v144, v33, v39, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
+; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v3, 0x7fff
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v106, 8, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v111, 8, v1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v32
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v146, v34, v36 :: v_dual_add_f32 v37, 0x40c00000, v37
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v31
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v36.l, v146.h
+; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v37, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v37
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v33.l, v148.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v35.l, v146.h
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v105, 24, v2
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v164, v34, v36, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v34, 0x40c00000, v38
-; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v35, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v107, 8, v2
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v36.l, v164.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v4, v39, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v34, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v37, 0x7fff
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v4, v39 :: v_dual_add_f32 v34, 0x40c00000, v38
; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v7
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v36, v3
; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff0000, v5
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v33.l, v133.h
+; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v34, 16, 1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v33, v144
-; GFX11-TRUE16-NEXT: v_add3_u32 v35, v37, v34, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 24, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v94, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v106, 8, v3
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v33, v135
; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v6, 16, 1
+; GFX11-TRUE16-NEXT: v_add3_u32 v35, v37, v34, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v6
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v108, 8, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v32
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v34
; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v6, 0x7fff
; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v36, 16, 1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v31
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v165, v33, v37, vcc_lo
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v93, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v95, 8, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v104, 8, v3
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v150, v33, v37, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v5, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v5
; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v36, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v34.l, v165.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v161, v35, v38, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v151, v35, v38, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v5, 0x7fff
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff0000, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v36
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v180, v33, v37 :: v_dual_add_f32 v35, 0x40c00000, v35
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v34.l, v150.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v166, v33, v37, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_lshlrev_b32 v8, 16, v8
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v37.l, v166.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v37.l, v180.h
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v35, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v8, 16, 1
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v6, v38, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v35
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v35, 0x7fff
; GFX11-TRUE16-NEXT: v_add3_u32 v6, v33, v8, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v33, 0x400000, v8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v179, v6, v33, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v177, v6, v33, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
-; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v34, v161
-; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v179.h
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v178, v5, v38 :: v_dual_add_f32 v33, 0x40c00000, v39
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v33, 0x40c00000, v39
+; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v34, v151
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v178, v5, v38, vcc_lo
; GFX11-TRUE16-NEXT: v_bfi_b32 v5, 0xffff, v37, v36
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v9
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v36, 16, v10
-; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v7 :: v_dual_lshlrev_b32 v36, 16, v10
; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v33, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33
+; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v177.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
+; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v7, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v7
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v36, 0x40c00000, v36
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v33
; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v33, 0x7fff
; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v178
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v47, v35, v37, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v36, 16, 1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v76, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v60, 24, v8
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v43, v35, v37, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v36
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v59, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 8, v8
; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v36, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v36
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 8, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v78, 8, v6
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v33, v34, v38, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v47.h
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v75, 24, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v88, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v91, 8, v5
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v44, v7, v37, vcc_lo
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v39
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX11-TRUE16-NEXT: v_bfe_u32 v36, v9, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v9
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, 0x400000, v7
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; GFX11-TRUE16-NEXT: v_add3_u32 v36, v36, v9, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v43.h
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v88, 8, v5
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v41, v7, v37, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v10, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v10
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v10, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v41, v35, v38 :: v_dual_lshlrev_b32 v10, 16, v12
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v12
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v44, v35, v38 :: v_dual_and_b32 v39, 0xffff0000, v9
; GFX11-TRUE16-NEXT: v_add_f32_e32 v37, 0x40c00000, v10
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v44.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v41.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_bfe_u32 v38, v37, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v50, 0x400000, v37
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v35, v41
-; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v35, v44
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_add3_u32 v38, v38, v37, 0x7fff
; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v37, 0x40c00000, v51
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v42, 24, v10
-; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v7, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v61, v38, v50 :: v_dual_add_f32 v12, 0x40c00000, v12
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v47, 8, v10
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v59, v38, v50, vcc_lo
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_lshlrev_b32 v7, 16, v9
; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v14
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v61.h
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v37, 0x40c00000, v51 :: v_dual_lshlrev_b32 v14, 16, v14
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-TRUE16-NEXT: v_bfe_u32 v48, v12, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v52, 0x400000, v12
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
+; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, 0x400000, v7
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-TRUE16-NEXT: v_add3_u32 v48, v48, v12, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v59.h
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v73, v35, v49, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v39
; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v37, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v48, v48, v12, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_lshlrev_b32 v11, 16, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v56, 8, v10
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v57, v48, v52, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT: v_bfe_u32 v49, v14, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v7, v57
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v36, v39, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v62, v48, v52, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v36, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v7, v62
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add3_u32 v36, v36, v9, 0x7fff
; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v11
; GFX11-TRUE16-NEXT: v_add3_u32 v11, v35, v37, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, 0x400000, v37
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v180, 24, v12
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v36, v39, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v73.h
; GFX11-TRUE16-NEXT: v_bfe_u32 v39, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 8, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v12
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v11, v11, v35, vcc_lo
; GFX11-TRUE16-NEXT: v_add_f32_e32 v35, 0x40c00000, v38
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_add3_u32 v37, v39, v7, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v7
; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff0000, v13
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT: v_bfe_u32 v48, v35, 16, 1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v36, v9
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v39
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v89, v37, v38, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v39 :: v_dual_cndmask_b32 v92, v37, v38
; GFX11-TRUE16-NEXT: v_add3_u32 v37, v48, v35, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v35
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v35, v35
@@ -160898,18 +160390,18 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, 0x400000, v14
; GFX11-TRUE16-NEXT: v_bfe_u32 v49, v7, 16, 1
; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v77, v37, v38, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v89, v37, v38, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, 0x400000, v7
; GFX11-TRUE16-NEXT: v_add3_u32 v14, v49, v7, 0x7fff
; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff0000, v16
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v78, v39, v48, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v91, v39, v48, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT: v_bfe_u32 v39, v13, 16, 1
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v78.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v73.h
; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v7, v14, v35 :: v_dual_add_f32 v14, 0x40c00000, v37
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v37, 16, v15
; GFX11-TRUE16-NEXT: v_add3_u32 v35, v39, v13, 0x7fff
@@ -160919,7 +160411,7 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_add_f32_e32 v37, 0x40c00000, v37
; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, 0x400000, v16
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v104, v35, v39, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v108, v35, v39, vcc_lo
; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v16, 0x7fff
; GFX11-TRUE16-NEXT: v_bfe_u32 v39, v37, 16, 1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
@@ -160927,405 +160419,366 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
; GFX11-TRUE16-NEXT: v_or_b32_e32 v51, 0x400000, v37
; GFX11-TRUE16-NEXT: v_add3_u32 v39, v39, v37, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v93, v13, v49, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v107, v13, v49, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX11-TRUE16-NEXT: v_add3_u32 v35, v48, v14, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, 0x400000, v14
; GFX11-TRUE16-NEXT: v_bfe_u32 v50, v15, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v15
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v95, v39, v51, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v110, v39, v51, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v104.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v108.h
; GFX11-TRUE16-NEXT: v_add3_u32 v13, v50, v15, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v89.h
-; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v38, v77
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v92, v35, v48, vcc_lo
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v91.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v92.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v109, v35, v48, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v95.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v93.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v110.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v107.h
+; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v38, v89
; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v39, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v14
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v13, v13, v16, vcc_lo
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v14
-; GFX11-TRUE16-NEXT: v_bfi_b32 v16, 0xffff, v35, v92
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18]
+; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v36, v9
+; GFX11-TRUE16-NEXT: v_bfi_b32 v16, 0xffff, v35, v109
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[23:24]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[11:12]
; GFX11-TRUE16-NEXT: v_bfi_b32 v15, 0xffff, v15, v13
; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v37, v7
; GFX11-TRUE16-NEXT: v_bfi_b32 v7, 0xffff, v34, v33
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[7:8]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[5:6]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[3:4]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[7:8]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[5:6]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[15:16]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[21:22]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18]
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v76, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v40, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v57, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v74, 8, v7
; GFX11-TRUE16-NEXT: .LBB90_4: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v108.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v131.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v111.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v133.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v107.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v68.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v132.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v2.h, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v129.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v1.l, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v128.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v6.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v106.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v164.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v1.h, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v105.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v2.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v94.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v91.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v148.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v67.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v180.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v90.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v144.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v14
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v4.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v165.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v88.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v5.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v47.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v76.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v58.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v14
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v75.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v161.h
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v6.h, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v179.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v72.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v6.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v105.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v69.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v2.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v146.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v104.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v95.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v2
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v135.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v3.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v133.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v93.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v68.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v5.l, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v166.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v88.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v5.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v6, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v78.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v150.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v151.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v76.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v66, v6, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v43.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v7.l, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v74.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v65.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v8.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v73.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v178.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v59.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v12, v14
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v56.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v8.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v44.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v43.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v89.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v41.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v42.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v16, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v10.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v61.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v183.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v16, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v11.l, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v104.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v176.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v166.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v16, v14
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v167.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v57.h
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v78.h
-; GFX11-TRUE16-NEXT: s_clause 0x1
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v12.l, v14.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v49.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v12.h, v15.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v16, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v160.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v77.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v95.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v150.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v64, v18, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v93.h
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v149.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v65, v18, v14
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v71.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v79.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v92.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v134.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v13.h, v15.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v66, v18, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v70.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v74.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v67, v18, v14
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v46.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v15.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v15.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v84.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v63.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v15.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v14
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v55.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v62.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v81.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v60.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v50.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v13.h, v16.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v20, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v86.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v17.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v80.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v45.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v19, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v18.h
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v13.h, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v83.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v40.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v18.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v19, v14
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v38.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v97.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v182.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v19.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v82.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v181.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v13.h, v19.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v22, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v87.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v23.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v177.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v37.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v22, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v24
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v147.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v13.h, v21.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v21.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v101.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v163.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v21.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v22, v14
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v85.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v162.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v98.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v151.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v36.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.h, v22.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v26, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v28
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v112.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v23.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v145.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v25, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v26
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v24.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.h, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v100.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v135.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v24.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v25, v14
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v35.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v113.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v131.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v26
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v25.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v99.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.l, 8, v130.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v13.h, v25.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v28, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v30
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v103.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v26.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v129.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v67.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v62.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v67, v6, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v177.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v7.l, v7.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v63.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v178.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v60.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v180.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v68, v6, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v73.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v7.l, v7.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v57.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v64.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v69, v6, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v41.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.l, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v47.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v44.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v45.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v6, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v92.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.l, v8.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v6.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v40.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.l, v10.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v6, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v59.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v182.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v50.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v6, v9
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v89.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v108.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v176.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.l, v11.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v6, v10
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v163.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v91.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v165.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v6, v11
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v110.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v161.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v38.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v109.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v6, v12
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v147.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.l, v13.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v107.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v149.l
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.l, v14.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v6, v13
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v82.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v94.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v65.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v6, v14
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v79.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.l, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v80.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v90.l
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.l, v16.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v6, v15
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v85.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v77.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v84.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v6, v16
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v72.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.l, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v83.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v75.l
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.l, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v6, v17
+; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v96.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v61.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v51.l
+; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v87.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v6, v18
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v56.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.l, v19.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v86.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v58.l
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.l, v20.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v6, v19
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v23.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v99.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v46.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v98.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v6, v20
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v183.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.l, v21.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v97.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v42.l
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.l, v22.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v6, v21
+; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v102.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v181.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v36.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v101.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v6, v22
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v167.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.l, v23.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v100.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v179.l
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.l, v24.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v6, v23
+; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v113.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v164.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v112.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v6, v24
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.l, v25.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v103.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v162.l
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.l, v26.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v6, v25
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v29.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v116.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v148.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v34.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v28, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff, v30
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v13.h, v27.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v27.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v116.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v128.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v27.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v28, v14
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v102.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.l, 8, v119.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v13.h
-; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff, v29
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v115.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v118.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v28.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v115.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v6, v26
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v144.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.l, v27.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v114.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v145.l
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.l, v28.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v6, v27
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v13.h, v28.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v32, v14
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff, v34
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v114.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff, v30
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v31, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.l, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v119.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v134.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v33.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v118.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v6, v28
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v130.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.l, v29.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v117.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v132.l
+; GFX11-TRUE16-NEXT: s_clause 0x1
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[66:69], off offset:16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v6, v29
+; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v30.l, v30.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v6.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v30, v14
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v6, v5
; GFX11-TRUE16-NEXT: s_clause 0x5
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
-; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[64:67], off offset:48
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[7:10], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[11:14], off offset:48
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[15:18], off offset:64
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[19:22], off offset:80
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[23:26], off offset:96
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[27:30], off offset:112
; GFX11-TRUE16-NEXT: s_clause 0x1f
-; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:12
-; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:16
-; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:20
-; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:24
-; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:32
-; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:40
-; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:48
-; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:52
-; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:56
-; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:60
-; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:64
-; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:72
-; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:80
-; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:84
-; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:92
-; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:100
-; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:108
-; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:116
-; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:124
-; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:128
-; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:132
-; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:136
-; GFX11-TRUE16-NEXT: s_clause 0x4
-; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:140
-; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:144
-; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:148
-; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:152
-; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:156
+; GFX11-TRUE16-NEXT: scratch_load_b32 v111, off, s32 offset:12
+; GFX11-TRUE16-NEXT: scratch_load_b32 v110, off, s32 offset:16
+; GFX11-TRUE16-NEXT: scratch_load_b32 v109, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:24
+; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:32
+; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:40
+; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:44
+; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:48
+; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:52
+; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:56
+; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:64
+; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:72
+; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:76
+; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:80
+; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:84
+; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:88
+; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:92
+; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:96
+; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:104
+; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:108
+; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:112
+; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:120
+; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:124
+; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:128
+; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:136
+; GFX11-TRUE16-NEXT: s_clause 0x7
+; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:140
+; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:144
+; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:148
+; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:152
+; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:156
+; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:160
+; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:164
+; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:168
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -185302,69 +184755,69 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_b32 v33, off, s32 offset:8
; GFX11-TRUE16-NEXT: scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr166_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr165_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr164_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
@@ -185375,69 +184828,69 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[7:8]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[5:6]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[7:8]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[11:12]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[3:4]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[5:6]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v27
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v21
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[15:16]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[1:2]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[23:24]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[21:22]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18]
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v17
; GFX11-TRUE16-NEXT: .LBB94_2: ; %Flow
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB94_4
@@ -185446,405 +184899,364 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_pk_add_f16 v32, 0x200, v32 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
-; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[7:8]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[5:6]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[7:8]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[11:12]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[3:4]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[5:6]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[15:16]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[23:24]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[21:22]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18]
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v27
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v21
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v17
; GFX11-TRUE16-NEXT: .LBB94_4: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v163.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v166.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v165.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v164.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v163.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v69.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v1.h, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v68, 0xffff, v68
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v54, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v67.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v68, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v67
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v4.l, v4.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v54, v51
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v68.l
+; 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_or_b32_e32 v4, v39, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v149.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v147.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v67
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v67.l
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v133.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v65.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v131.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v65
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v8.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v119.l
; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v113.l
; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v65
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v118.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v116.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v99.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v85.l
; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v103.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v150.l
; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v49.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v51
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v97.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v49, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v148.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l
; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v49, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v84.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v48, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v145.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l
; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v116.l
; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v134.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v114.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l
; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v18.l, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v19.l, v19.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v20.l
-; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v128.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v102.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l
; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v96.l
; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v22.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v21.l, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v114.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v83.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v22.l, v22.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v113.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l
; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v81.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v100.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v24.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v51
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v25.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v86.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v27.l, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v82.l
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v28.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v81.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v71.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v30.l, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v51
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v31.l, v31.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l
-; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v55.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v32.l, v32.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v51
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32
; GFX11-TRUE16-NEXT: s_clause 0x5
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
@@ -208055,69 +207467,69 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_b32 v33, off, s32 offset:8
; GFX11-TRUE16-NEXT: scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr166_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr165_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr164_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
@@ -208128,69 +207540,69 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[7:8]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[5:6]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[7:8]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[11:12]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[3:4]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[5:6]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v27
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v21
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[15:16]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[1:2]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[23:24]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[21:22]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18]
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v17
; GFX11-TRUE16-NEXT: .LBB98_2: ; %Flow
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_4
@@ -208199,405 +207611,364 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
-; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[7:8]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[5:6]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[7:8]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[11:12]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[3:4]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[15:16]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[5:6]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[15:16]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[13:14]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[11:12]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[23:24]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[21:22]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[17:18]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[23:24]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[21:22]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18]
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 24, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v32
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v31
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v30
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v29
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v28
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v27
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v26
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v25
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v24
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 8, v23
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v21
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v20
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v18
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v17
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 8, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 8, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v1
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v31
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v30
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v29
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v28
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v27
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v26
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v25
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 24, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v23
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v21
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 24, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v19
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 24, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v18
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v17
; GFX11-TRUE16-NEXT: .LBB98_4: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v163.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v162.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v166.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v165.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v164.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v163.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v69.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v161.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v1.h, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v160.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v68, 0xffff, v68
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v54, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v67.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v68, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v149.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v67
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v4.l, v4.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v148.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v54, v51
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v68.l
+; 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_or_b32_e32 v4, v39, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v149.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v147.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v5.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v5.h, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v67
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v67.l
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v133.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v146.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v65.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v66
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v131.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v65
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v8.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v119.l
; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v113.l
; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v54, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v64.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v65
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v118.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v52.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v116.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v54, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v64
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v11.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v11.h, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v99.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v85.l
; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v103.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v150.l
; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v52, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v49.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v54
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v51
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v97.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v49, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v14.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v148.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l
; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l
; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v49, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v48, 0xffff, v52
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v84.l
-; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v39
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v48, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.h
-; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v145.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v17.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l
-; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v17.h, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l
; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l
+; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v116.l
; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v134.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l
+; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v114.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l
; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v18.l, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v19.l, v19.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v20.l
-; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v128.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v39, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v20.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l
-; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v20.h, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v102.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l
+; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l
; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v96.l
; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v22.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v48
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.h
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v21.l, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v114.l
-; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v83.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v38, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v22.l, v22.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.h
-; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v113.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v38, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v39
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v23.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v23.h, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l
; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v81.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v37, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v33.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v100.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v37, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v24.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v38
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v37, v51
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v25.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v26.l
-; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v86.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v36, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v36, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v26.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l
-; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v26.h, v27.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v36, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v37
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.h
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v27.l, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v82.l
-; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v34.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v35, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v35, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v28.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.l, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.l, 8, v81.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v35, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v36
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v29.l, v33.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v29.h, v30.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v34, v51
-; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v33.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v30.l
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v71.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v31.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v30.l, v30.h
-; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v32.l, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v35
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v34, v51
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v31.l, v31.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v32.l
-; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v55.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v33, v51
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v33, 0xffff, v34
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v51.h, v32.l, v32.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v33, v51
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31
+; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32
; GFX11-TRUE16-NEXT: s_clause 0x5
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll
index 3e96ab1d597d6..21ec3ee1996a6 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll
@@ -4118,19 +4118,19 @@ define <4 x i32> @bitcast_v16i8_to_v4i32(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16
@@ -4144,103 +4144,95 @@ define <4 x i32> @bitcast_v16i8_to_v4i32(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2
; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -8592,19 +8584,19 @@ define <4 x float> @bitcast_v16i8_to_v4f32(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16
@@ -8618,103 +8610,95 @@ define <4 x float> @bitcast_v16i8_to_v4f32(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2
; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -12682,19 +12666,19 @@ define <2 x i64> @bitcast_v16i8_to_v2i64(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16
@@ -12708,103 +12692,95 @@ define <2 x i64> @bitcast_v16i8_to_v2i64(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2
; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -16382,19 +16358,19 @@ define <2 x double> @bitcast_v16i8_to_v2f64(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16
@@ -16408,103 +16384,95 @@ define <2 x double> @bitcast_v16i8_to_v2f64(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2
; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -19811,19 +19779,19 @@ define <8 x i16> @bitcast_v16i8_to_v8i16(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16
@@ -19837,103 +19805,95 @@ define <8 x i16> @bitcast_v16i8_to_v8i16(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB98_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_2
; GFX11-TRUE16-NEXT: .LBB98_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -22725,19 +22685,19 @@ define <8 x half> @bitcast_v16i8_to_v8f16(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16
@@ -22751,103 +22711,95 @@ define <8 x half> @bitcast_v16i8_to_v8f16(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB106_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB106_2
; GFX11-TRUE16-NEXT: .LBB106_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -24944,19 +24896,19 @@ define <8 x bfloat> @bitcast_v16i8_to_v8bf16(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16
@@ -24970,103 +24922,95 @@ define <8 x bfloat> @bitcast_v16i8_to_v8bf16(<16 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB110_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v8.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v7.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v1.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v3.l, v4.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v9
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB110_2
; GFX11-TRUE16-NEXT: .LBB110_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v9.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v6.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v11
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v4, v11
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll
index f8ffaa456c2b3..38302a75fe26d 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll
@@ -6296,32 +6296,31 @@ define <8 x i32> @bitcast_v32i8_to_v8i32(<32 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v22.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v23.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v21.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.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
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v31.l
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB26_3
@@ -6333,194 +6332,175 @@ define <8 x i32> @bitcast_v32i8_to_v8i32(<32 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v0.h, v15.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v12.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v1.h, v14.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v2.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v11.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v3.l, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v12, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v4.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v17.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v0.l, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v19.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v1.h, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v2.l, v14.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v3.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v13.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v10.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v4.l, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_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 v4, v7, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v5.l, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v9.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v6.l, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v7.l, v8.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v5.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v6.l, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v7.l, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v19
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2
; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v17.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v16.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v19.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v18.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v17.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v15.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v15.h, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v16.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v14.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v15.l, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v14.h, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v16.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v13.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v12.l, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v12.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v18.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.h, v3.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v12.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v13.l, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v11.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v11.h, v3.h
+; 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: v_mov_b16_e32 v6.l, v4.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v10.h, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v21
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v11.l, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v21
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v9.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v10.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v9.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.h, v6.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v21
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v21
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v8, v21
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v8.h, v6.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -13335,32 +13315,31 @@ define <8 x float> @bitcast_v32i8_to_v8f32(<32 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v22.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v23.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v21.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.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
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v31.l
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB50_3
@@ -13372,194 +13351,175 @@ define <8 x float> @bitcast_v32i8_to_v8f32(<32 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v0.h, v15.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v12.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v1.h, v14.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v2.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v11.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v3.l, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v12, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v4.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v17.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v0.l, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v19.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v1.h, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v2.l, v14.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v3.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v13.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v10.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v4.l, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_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 v4, v7, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v5.l, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v9.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v6.l, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v7.l, v8.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v5.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v6.l, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v7.l, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v19
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2
; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v17.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v16.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v19.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v18.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v17.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v15.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v15.h, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v16.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v14.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v15.l, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v14.h, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v16.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v13.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v12.l, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v12.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v18.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.h, v3.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v12.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v13.l, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v11.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v11.h, v3.h
+; 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: v_mov_b16_e32 v6.l, v4.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v10.h, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v21
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v11.l, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v21
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v9.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v10.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v9.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.h, v6.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v21
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v21
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v8, v21
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v8.h, v6.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -19892,32 +19852,31 @@ define <4 x i64> @bitcast_v32i8_to_v4i64(<32 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v22.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v23.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v21.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.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
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v31.l
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB70_3
@@ -19929,194 +19888,175 @@ define <4 x i64> @bitcast_v32i8_to_v4i64(<32 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v0.h, v15.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v12.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v1.h, v14.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v2.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v11.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v3.l, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v12, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v4.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v17.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v0.l, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v19.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v1.h, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v2.l, v14.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v3.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v13.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v10.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v4.l, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_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 v4, v7, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v5.l, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v9.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v6.l, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v7.l, v8.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v5.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v6.l, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v7.l, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v19
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2
; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v17.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v16.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v19.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v18.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v17.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v15.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v15.h, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v16.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v14.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v15.l, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v14.h, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v16.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v13.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v12.l, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v12.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v18.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.h, v3.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v12.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v13.l, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v11.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v11.h, v3.h
+; 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: v_mov_b16_e32 v6.l, v4.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v10.h, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v21
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v11.l, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v21
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v9.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v10.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v9.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.h, v6.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v21
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v21
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v8, v21
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v8.h, v6.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -25939,32 +25879,31 @@ define <4 x double> @bitcast_v32i8_to_v4f64(<32 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v12.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v22.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v21.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v23.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v21.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.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
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v31.l
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB86_3
@@ -25976,194 +25915,175 @@ define <4 x double> @bitcast_v32i8_to_v4f64(<32 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v0.h, v15.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v12.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v1.h, v14.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v2.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v11.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v3.l, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v12, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v4.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v17.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v0.l, v19.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v19.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v1.h, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v2.l, v14.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v3.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v13.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v10.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v4.l, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_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 v4, v7, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v5.l, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v9.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v6.l, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v7.l, v8.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v5.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v6.l, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v7.l, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v19
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2
; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v17.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v16.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v19.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v18.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v17.h, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v15.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v15.h, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v16.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v14.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v15.l, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v14.h, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v16.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v13.l, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v12.l, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v12.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v18.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.h, v3.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v21
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v12.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v13.l, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v11.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v11.h, v3.h
+; 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: v_mov_b16_e32 v6.l, v4.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v10.h, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v21
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v11.l, v4.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v24.l, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v21
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v9.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v10.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v9.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.h, v6.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v21
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v21
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v8, v21
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v8.h, v6.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll
index 0cefbc1c2dee5..436b1a038b274 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: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_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: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 v30, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1
; GFX11-TRUE16-NEXT: .LBB12_4: ; %end
@@ -3047,105 +3047,93 @@ 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 v2.l, 0xff, v2.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_lshlrev_b16 v12.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, 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_lshlrev_b16 v12.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
; 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_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
; 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_lshlrev_b16 v12.h, 8, v26.l
+; 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_or_b32_e32 v3, v15, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
; 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_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
-; 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_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.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_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_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: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
+; 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_b32_e32 v11, v15, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
; 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[10:11], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v10i32_to_v40i8:
@@ -5038,48 +5026,49 @@ 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_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 v31, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, 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 v35.h, v17.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 v17.l, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 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 v24.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v5.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 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: v_lshlrev_b16 v23.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v35.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v30.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v29.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v27.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.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 v11.l, 8, v33.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v33.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v33.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(6)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v34.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.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: v_lshlrev_b16 v10.l, 8, v34.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v35.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 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
@@ -5092,245 +5081,217 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB14_3: ; %cmp.false
; 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 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_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: v_mov_b16_e32 v27.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v0.l, v24.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v25.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v23.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v19.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v27, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v1.h, v23.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v27, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v2.l, v20.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v17.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v27, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v3.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v14.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v27, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v4.l, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v27, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v5.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v27, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v6.l, v13.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v27.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_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: $vgpr17_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: $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: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_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: $vgpr23_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; 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: $vgpr18_hi16
; 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: $vgpr15_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: $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: $vgpr13_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16
; 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: $vgpr13_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v10.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v27, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v7.l, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v27.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v27, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v8.l, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v27.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v27, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v9.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v27.h
; 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: v_or_b32_e32 v9, v27, v9
; 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 v2.l, v22.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v23.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v22.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, 0
; 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 v2.l, 0xff, v2.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_add_nc_u16 v3.l, v20.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.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_mov_b16_e32 v27.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v24.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v25.l, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v19.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v19.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v25, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v16.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v20.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v21.l, v1.h
; 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_or_b32_e32 v1, v25, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v25.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v18.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v25, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v15.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v15.h, v3.h
; 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_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_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_add_nc_u16 v5.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v25, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v14.l, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v14.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v13.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v13.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; 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: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v25, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v12.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v12.h, v6.h
; 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_add_nc_u16 v8.l, v32.h, 3
; 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 v6, v25, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.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_add_nc_u16 v8.h, v31.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v11.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v7.h
; 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: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v25, v12
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.h, v8.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v25, v9
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -9951,20 +9912,20 @@ 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: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_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: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
@@ -9980,17 +9941,17 @@ 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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 v30, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1
; GFX11-TRUE16-NEXT: .LBB32_2: ; %Flow
@@ -10010,17 +9971,17 @@ 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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 v30, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1
; GFX11-TRUE16-NEXT: .LBB32_4: ; %end
@@ -10028,105 +9989,93 @@ 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 v2.l, 0xff, v2.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_lshlrev_b16 v12.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, 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_lshlrev_b16 v12.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
; 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_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
; 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_lshlrev_b16 v12.h, 8, v26.l
+; 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_or_b32_e32 v3, v15, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
; 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_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
-; 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_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.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_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_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: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
+; 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_b32_e32 v11, v15, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
; 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[10:11], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v10f32_to_v40i8:
@@ -12037,48 +11986,49 @@ 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_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 v31, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, 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 v35.h, v17.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 v17.l, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 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 v24.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v5.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 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: v_lshlrev_b16 v23.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v35.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v30.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v29.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v27.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.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 v11.l, 8, v33.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v33.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v33.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(6)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v34.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.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: v_lshlrev_b16 v10.l, 8, v34.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v35.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 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
@@ -12091,245 +12041,217 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB34_3: ; %cmp.false
; 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 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_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: v_mov_b16_e32 v27.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v22.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.h
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v0.l, v24.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v25.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v23.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v19.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v27, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v1.h, v23.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v27, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v2.l, v20.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v17.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v15.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v27, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v3.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v14.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v27, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v4.l, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v27, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v5.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v27, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v6.l, v13.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v27.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_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: $vgpr17_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: $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: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_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: $vgpr23_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; 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: $vgpr18_hi16
; 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: $vgpr15_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: $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: $vgpr13_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16
; 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: $vgpr13_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v10.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v27, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v7.l, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v27.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v27, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v8.l, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v27.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v27, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v9.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v27.h
; 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: v_or_b32_e32 v9, v27, v9
; 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 v2.l, v22.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v23.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v22.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, 0
; 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 v2.l, 0xff, v2.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_add_nc_u16 v3.l, v20.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.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_mov_b16_e32 v27.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v24.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v25.l, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v19.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v19.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v25, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v16.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v20.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v21.l, v1.h
; 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_or_b32_e32 v1, v25, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v25.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v18.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v25, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v15.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v15.h, v3.h
; 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_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_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_add_nc_u16 v5.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v25, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v14.l, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v14.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v13.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v13.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; 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: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v25, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v12.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v12.h, v6.h
; 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_add_nc_u16 v8.l, v32.h, 3
; 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 v6, v25, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.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_add_nc_u16 v8.h, v31.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v11.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v7.h
; 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: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v25, v12
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.h, v8.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v25, v9
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -16358,20 +16280,20 @@ 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: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_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: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
@@ -16387,17 +16309,17 @@ 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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 v30, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1
; GFX11-TRUE16-NEXT: .LBB48_2: ; %Flow
@@ -16421,17 +16343,17 @@ 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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 v30, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1
; GFX11-TRUE16-NEXT: .LBB48_4: ; %end
@@ -16439,105 +16361,93 @@ 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 v2.l, 0xff, v2.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_lshlrev_b16 v12.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, 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_lshlrev_b16 v12.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
; 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_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
; 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_lshlrev_b16 v12.h, 8, v26.l
+; 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_or_b32_e32 v3, v15, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
; 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_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
-; 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_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.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_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_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: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
+; 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_b32_e32 v11, v15, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
; 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[10:11], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v20i16_to_v40i8:
@@ -22479,20 +22389,20 @@ 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: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_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: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
@@ -22508,17 +22418,17 @@ 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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 v30, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1
; GFX11-TRUE16-NEXT: .LBB60_2: ; %Flow
@@ -22542,17 +22452,17 @@ 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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 v30, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1
; GFX11-TRUE16-NEXT: .LBB60_4: ; %end
@@ -22560,105 +22470,93 @@ 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 v2.l, 0xff, v2.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_lshlrev_b16 v12.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, 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_lshlrev_b16 v12.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
; 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_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
; 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_lshlrev_b16 v12.h, 8, v26.l
+; 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_or_b32_e32 v3, v15, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
; 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_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
-; 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_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.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_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_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: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
+; 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_b32_e32 v11, v15, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
; 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[10:11], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v20f16_to_v40i8:
@@ -28859,50 +28757,51 @@ 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_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 v31, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, 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, v21.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 v21.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 v25.l, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.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 v34.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v5.l
-; 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: v_lshlrev_b16 v33.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v21.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.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 v17.l, 8, v36.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v36.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v36.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(6)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.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: v_lshlrev_b16 v16.l, 8, v37.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v38.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 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
@@ -28915,245 +28814,216 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB72_3: ; %cmp.false
; 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 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_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_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: v_mov_b16_e32 v10.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v0.l, v34.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v33.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v1.h, v33.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v2.l, v28.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v10.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v3.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v26.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v22.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v4.l, v23.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v5.l, v21.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v6.l, v19.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h
; 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: $vgpr33_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_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: $vgpr25_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; 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: $vgpr34_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
; 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: $vgpr33_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
-; 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: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_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: $vgpr17_lo16
-; 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: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v16.h
; 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: v_or_b32_e32 v6, v10, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.l, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v9.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
; 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: v_or_b32_e32 v9, v10, v9
; 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 v2.l, v30.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v33.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v30.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, 0
; 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 v2.l, 0xff, v2.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_add_nc_u16 v3.l, v28.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.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_mov_b16_e32 v11.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v34.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v34.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v27.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v25.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v24.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v28.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v29.l, v1.h
; 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_or_b32_e32 v1, v10, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v26.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v21.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v23.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v23.h, v3.h
; 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_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_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_add_nc_u16 v5.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v21.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v22.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v19.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v19.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; 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: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v18.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v18.h, v6.h
; 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_add_nc_u16 v8.l, v32.h, 3
; 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 v6, v10, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.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_add_nc_u16 v8.h, v31.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v17.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v17.h, v7.h
; 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: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v16.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v16.h, v8.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v9
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -30908,20 +30778,20 @@ 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: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_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: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
@@ -30937,17 +30807,17 @@ 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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 v30, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1
; GFX11-TRUE16-NEXT: .LBB74_2: ; %Flow
@@ -30966,17 +30836,17 @@ 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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 v30, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1
; GFX11-TRUE16-NEXT: .LBB74_4: ; %end
@@ -30984,105 +30854,93 @@ 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 v2.l, 0xff, v2.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_lshlrev_b16 v12.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, 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_lshlrev_b16 v12.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
; 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_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
; 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_lshlrev_b16 v12.h, 8, v26.l
+; 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_or_b32_e32 v3, v15, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
; 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_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
-; 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_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.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_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_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: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
+; 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_b32_e32 v11, v15, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
; 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[10:11], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v5f64_to_v40i8:
@@ -33010,50 +32868,51 @@ 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_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 v31, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, 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, v21.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 v21.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 v25.l, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.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 v34.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v5.l
-; 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: v_lshlrev_b16 v33.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v21.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v39.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.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 v17.l, 8, v36.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v36.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v36.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(6)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.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: v_lshlrev_b16 v16.l, 8, v37.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v38.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 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
@@ -33066,245 +32925,216 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB76_3: ; %cmp.false
; 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 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_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_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: v_mov_b16_e32 v10.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v0.l, v34.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v34.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v33.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v1.h, v33.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v20.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v2.l, v28.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v10.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v3.l, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v26.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v22.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v4.l, v23.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v5.l, v21.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v6.l, v19.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h
; 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: $vgpr33_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_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: $vgpr25_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; 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: $vgpr34_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
; 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: $vgpr33_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
-; 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: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_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: $vgpr17_lo16
-; 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: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v16.h
; 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: v_or_b32_e32 v6, v10, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.l, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v9.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
; 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: v_or_b32_e32 v9, v10, v9
; 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 v2.l, v30.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v33.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v30.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, 0
; 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 v2.l, 0xff, v2.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_add_nc_u16 v3.l, v28.h, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.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_mov_b16_e32 v11.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v34.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v34.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v27.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v25.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v24.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v28.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v29.l, v1.h
; 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_or_b32_e32 v1, v10, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v26.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v21.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v23.l, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v23.h, v3.h
; 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_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_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_add_nc_u16 v5.l, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v21.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v22.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v19.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v19.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; 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: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v18.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v18.h, v6.h
; 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_add_nc_u16 v8.l, v32.h, 3
; 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 v6, v10, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.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_add_nc_u16 v8.h, v31.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v17.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v17.h, v7.h
; 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: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v16.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v16.h, v8.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v9
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -35074,20 +34904,20 @@ 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: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_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: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
@@ -35103,17 +34933,17 @@ 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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 v30, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1
; GFX11-TRUE16-NEXT: .LBB78_2: ; %Flow
@@ -35140,17 +34970,17 @@ 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 v20, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v19, 24, v8
+; 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 v21, 8, v8
-; 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 v22, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v23, 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 v29, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v2
+; 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 v30, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v16, 8, v1
; GFX11-TRUE16-NEXT: .LBB78_4: ; %end
@@ -35158,105 +34988,93 @@ 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 v2.l, 0xff, v2.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_lshlrev_b16 v12.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, 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_lshlrev_b16 v12.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l
; 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_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h
; 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_lshlrev_b16 v12.h, 8, v26.l
+; 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_or_b32_e32 v3, v15, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
; 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_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l
-; 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_lshlrev_b16 v12.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.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_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_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: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h
+; 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_b32_e32 v11, v15, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9
; 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[10:11], off offset:32
+; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v5i64_to_v40i8:
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll
index 48c9b8775a474..8e30ee659a260 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll
@@ -2257,8 +2257,8 @@ define i32 @bitcast_v4i8_to_i32(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4
@@ -2273,19 +2273,17 @@ define i32 @bitcast_v4i8_to_i32(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB22_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB22_2
; GFX11-TRUE16-NEXT: .LBB22_4: ; %cmp.true
@@ -2295,16 +2293,14 @@ define i32 @bitcast_v4i8_to_i32(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h
; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -4506,8 +4502,8 @@ define float @bitcast_v4i8_to_f32(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4
@@ -4522,19 +4518,17 @@ define float @bitcast_v4i8_to_f32(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB42_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB42_2
; GFX11-TRUE16-NEXT: .LBB42_4: ; %cmp.true
@@ -4544,16 +4538,14 @@ define float @bitcast_v4i8_to_f32(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h
; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -6467,8 +6459,8 @@ define <2 x i16> @bitcast_v4i8_to_v2i16(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4
@@ -6483,19 +6475,17 @@ define <2 x i16> @bitcast_v4i8_to_v2i16(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB58_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB58_2
; GFX11-TRUE16-NEXT: .LBB58_4: ; %cmp.true
@@ -6505,16 +6495,14 @@ define <2 x i16> @bitcast_v4i8_to_v2i16(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h
; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -8116,8 +8104,8 @@ define <2 x half> @bitcast_v4i8_to_v2f16(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4
@@ -8132,19 +8120,17 @@ define <2 x half> @bitcast_v4i8_to_v2f16(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2
; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true
@@ -8154,16 +8140,14 @@ define <2 x half> @bitcast_v4i8_to_v2f16(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h
; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -9479,8 +9463,8 @@ define <2 x bfloat> @bitcast_v4i8_to_v2bf16(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4
@@ -9495,19 +9479,17 @@ define <2 x bfloat> @bitcast_v4i8_to_v2bf16(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB78_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB78_2
; GFX11-TRUE16-NEXT: .LBB78_4: ; %cmp.true
@@ -9517,16 +9499,14 @@ define <2 x bfloat> @bitcast_v4i8_to_v2bf16(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h
; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -10193,8 +10173,8 @@ define <1 x i32> @bitcast_v4i8_to_v1i32(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v4
@@ -10209,19 +10189,17 @@ define <1 x i32> @bitcast_v4i8_to_v1i32(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB82_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB82_2
; GFX11-TRUE16-NEXT: .LBB82_4: ; %cmp.true
@@ -10231,16 +10209,14 @@ define <1 x i32> @bitcast_v4i8_to_v1i32(<4 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h
; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
index 5aac06a7f3a2b..35d135b123969 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
@@ -8768,32 +8768,32 @@ define <64 x i8> @bitcast_v16i32_to_v64i8(<16 x i32> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
@@ -8812,26 +8812,26 @@ define <64 x i8> @bitcast_v16i32_to_v64i8(<16 x i32> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB24_2: ; %Flow
@@ -8864,26 +8864,26 @@ define <64 x i8> @bitcast_v16i32_to_v64i8(<16 x i32> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB24_4: ; %end
@@ -8891,156 +8891,135 @@ define <64 x i8> @bitcast_v16i32_to_v64i8(<16 x i32> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l
+; 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_or_b32_e32 v3, v24, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l
; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16
; GFX11-TRUE16-NEXT: s_clause 0x3
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
@@ -12470,15 +12449,15 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:128
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:124
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:116
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:108
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:108
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:100
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:92
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:92
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v82, off, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_load_b32 v81, off, s32 offset:132
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:8
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:16
@@ -12492,84 +12471,82 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:80
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:84
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:60
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:52
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:20
; GFX11-TRUE16-NEXT: s_clause 0x1
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:12
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:12
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, v27.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v25.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.h, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v27.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.l, 8, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v81.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v80.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v27.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v64.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v65.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v65.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(25)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v66.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v66.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(23)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v66.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v66.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(21)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v67.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(20)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v67.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v67.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(19)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v68.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v68.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v68.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v68.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v69.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v69.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(15)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(14)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v70.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v70.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(12)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v71.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v82
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v80.l
+; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v81
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB26_3
; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow
@@ -12581,384 +12558,338 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v54.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v51.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v53.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v53.l
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v0.h, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v50.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v51.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v48.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v49.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v1.h, v50.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v39.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v2.l, v49.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v3.l, v48.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v53.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v52.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v52.l
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v0.l, v54.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v55.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v54.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v49.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v64, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v1.h, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v51.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v64, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v2.l, v51.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v48.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v39.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v64, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v3.l, v50.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v30.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v64, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v4.l, v39.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v38.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v4.l, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v38.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v64, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v5.l, v29.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v37.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v37.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v64, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v6.l, v27.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v36.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v64, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v7.l, v25.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v64, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v8.l, v23.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v20.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v33.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v9.l, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v64, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v10.l, v21.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v31.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v5.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v38.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v6.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v37.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v7.l, v23.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v37.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v35.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v8.l, v22.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v36.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v22.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v35.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v9.l, v21.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v21.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
+; 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: $vgpr37_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v10.l, v20.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v34.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v20.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v11.l, v19.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v19.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v12.l, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v18.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v13.l, v17.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v14.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_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: $vgpr16_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v54
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v15.l, v16.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v64, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v11.l, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v16.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v64, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v12.l, v19.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v64.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v64, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v13.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v64.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v64, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v14.l, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v64.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v64, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v15.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v64.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v54
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v64, v15
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2
; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v54.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v51.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v53.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v53.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v52.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v52.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, 0
; 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 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v48.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v53.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v50.l, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v52.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v55.l, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v54.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v53.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v49.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v48.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v52, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v39.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v51.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v51.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v52.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v50.h, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v39.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v30.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v55
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v49.h, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v51.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v55
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v48.l, v3.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v27.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v49.l, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v27.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v52, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v52.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v50.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v50.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v52, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v24.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v39.h, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.l, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v24.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.l, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v29.h, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v25.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v29.l, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v28.h, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v52, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v29.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v30.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v52, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v27.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v27.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v26.h, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v52, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v6.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v38.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v25.h, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v24.l, v7.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v23.h, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v52, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v37.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v23.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v23.h, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v22.h, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v36.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v23.l, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v21.h, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v55
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v52, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.h, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v52, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v35.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v21.l, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v21.h, v9.h
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v20.h, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v21.l, v10.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v34.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.h, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v52, v12
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v10.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v20.h, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v52, v13
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v33.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v19.l, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v19.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v19.l, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v18.h, v12.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v55
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.l, 3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v32.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v14
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v18.h, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v15
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v17.l, v13.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v31.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v17.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.h, v14.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v13.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v55
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v14.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v55
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v15.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v55
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v52, v18
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.h, v14.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v52, v17
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v52, v15
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -23588,32 +23519,32 @@ define <64 x i8> @bitcast_v16f32_to_v64i8(<16 x float> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
@@ -23632,26 +23563,26 @@ define <64 x i8> @bitcast_v16f32_to_v64i8(<16 x float> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB48_2: ; %Flow
@@ -23676,26 +23607,26 @@ define <64 x i8> @bitcast_v16f32_to_v64i8(<16 x float> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB48_4: ; %end
@@ -23703,156 +23634,135 @@ define <64 x i8> @bitcast_v16f32_to_v64i8(<16 x float> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l
+; 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_or_b32_e32 v3, v24, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l
; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16
; GFX11-TRUE16-NEXT: s_clause 0x3
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
@@ -27413,15 +27323,15 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:128
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:124
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:116
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:108
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:108
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:100
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:92
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:92
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v82, off, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_load_b32 v81, off, s32 offset:132
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:8
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:16
@@ -27435,84 +27345,82 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:80
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:84
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:60
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:52
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:20
; GFX11-TRUE16-NEXT: s_clause 0x1
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:12
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:12
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, v27.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v25.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.h, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v27.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.l, 8, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v81.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v80.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v27.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v64.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v65.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v65.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(25)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v66.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v66.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(23)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v66.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v66.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(21)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v67.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(20)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v67.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v67.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(19)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v68.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v68.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v68.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v68.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v69.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v69.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(15)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(14)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v70.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v70.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(12)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v71.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v82
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v80.l
+; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v81
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB50_3
; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow
@@ -27524,384 +27432,338 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v54.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v51.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v53.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v53.l
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v0.h, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v50.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v51.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v48.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v49.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v1.h, v50.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v39.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v2.l, v49.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v3.l, v48.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v53.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v52.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v52.l
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v0.l, v54.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v55.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v54.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v49.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v64, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v1.h, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v51.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v64, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v2.l, v51.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v48.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v39.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v64, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v3.l, v50.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v30.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v64, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v4.l, v39.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v38.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v4.l, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v38.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v64, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v5.l, v29.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v37.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v37.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v64, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v6.l, v27.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v36.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v64, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v7.l, v25.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v64, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v8.l, v23.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v20.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v33.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v9.l, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v64, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v10.l, v21.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v31.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v5.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v38.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v6.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v37.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v7.l, v23.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v37.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v35.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v8.l, v22.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v36.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v22.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v35.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v9.l, v21.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v21.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
+; 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: $vgpr37_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v10.l, v20.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v34.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v20.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v11.l, v19.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v19.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v12.l, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v18.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v13.l, v17.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v14.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_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: $vgpr16_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v54
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v15.l, v16.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v64, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v11.l, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v16.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v64, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v12.l, v19.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v64.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v64, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v13.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v64.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v64, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v14.l, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v64.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v64, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v15.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v64.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v54
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v64, v15
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2
; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v54.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v51.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v53.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v53.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v52.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v52.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, 0
; 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 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v48.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v53.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v50.l, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v52.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v55.l, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v54.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v53.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v49.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v48.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v52, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v39.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v51.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v51.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v52.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v50.h, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v39.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v30.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v55
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v49.h, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v51.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v55
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v48.l, v3.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v27.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v49.l, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v27.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v52, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v52.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v50.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v50.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v52, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v24.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v39.h, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.l, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v24.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.l, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v29.h, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v25.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v29.l, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v28.h, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v52, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v29.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v30.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v52, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v27.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v27.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v26.h, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v52, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v6.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v38.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v25.h, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v24.l, v7.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v23.h, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v52, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v37.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v23.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v23.h, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v22.h, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v36.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v23.l, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v21.h, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v55
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v52, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.h, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v52, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v35.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v21.l, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v21.h, v9.h
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v20.h, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v21.l, v10.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v34.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.h, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v52, v12
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v10.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v20.h, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v52, v13
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v33.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v19.l, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v19.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v19.l, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v18.h, v12.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v55
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.l, 3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v32.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v14
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v18.h, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v15
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v17.l, v13.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v31.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v17.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.h, v14.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v13.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v55
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v14.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v55
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v15.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v55
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v52, v18
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.h, v14.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v52, v17
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v52, v15
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -37916,32 +37778,32 @@ define <64 x i8> @bitcast_v8i64_to_v64i8(<8 x i64> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
@@ -37960,26 +37822,26 @@ define <64 x i8> @bitcast_v8i64_to_v64i8(<8 x i64> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB68_2: ; %Flow
@@ -38017,26 +37879,26 @@ define <64 x i8> @bitcast_v8i64_to_v64i8(<8 x i64> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB68_4: ; %end
@@ -38044,156 +37906,135 @@ define <64 x i8> @bitcast_v8i64_to_v64i8(<8 x i64> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l
+; 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_or_b32_e32 v3, v24, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l
; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16
; GFX11-TRUE16-NEXT: s_clause 0x3
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
@@ -41628,15 +41469,15 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:128
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:124
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:116
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:108
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:108
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:100
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:92
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:92
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v82, off, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_load_b32 v81, off, s32 offset:132
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:8
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:16
@@ -41650,84 +41491,82 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:80
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:84
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:60
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:52
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:20
; GFX11-TRUE16-NEXT: s_clause 0x1
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:12
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:12
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, v27.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v25.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.h, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v27.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.l, 8, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v81.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v80.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v27.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v64.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v65.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v65.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(25)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v66.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v66.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(23)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v66.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v66.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(21)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v67.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(20)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v67.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v67.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(19)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v68.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v68.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v68.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v68.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v69.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v69.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(15)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(14)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v70.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v70.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(12)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v71.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v82
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v80.l
+; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v81
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB70_3
; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow
@@ -41739,384 +41578,338 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v54.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v51.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v53.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v53.l
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v0.h, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v50.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v51.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v48.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v49.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v1.h, v50.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v39.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v2.l, v49.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v3.l, v48.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v53.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v52.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v52.l
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v0.l, v54.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v55.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v54.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v49.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v64, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v1.h, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v51.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v64, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v2.l, v51.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v48.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v39.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v64, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v3.l, v50.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v30.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v64, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v4.l, v39.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v38.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v4.l, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v38.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v64, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v5.l, v29.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v37.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v37.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v64, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v6.l, v27.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v36.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v64, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v7.l, v25.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v64, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v8.l, v23.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v20.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v33.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v9.l, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v64, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v10.l, v21.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v31.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v5.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v38.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v6.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v37.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v7.l, v23.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v37.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v35.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v8.l, v22.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v36.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v22.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v35.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v9.l, v21.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v21.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
+; 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: $vgpr37_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v10.l, v20.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v34.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v20.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v11.l, v19.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v19.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v12.l, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v18.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v13.l, v17.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v14.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_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: $vgpr16_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v54
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v15.l, v16.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v64, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v11.l, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v16.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v64, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v12.l, v19.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v64.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v64, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v13.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v64.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v64, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v14.l, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v64.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v64, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v15.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v64.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v54
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v64, v15
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2
; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v54.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v51.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v53.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v53.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v52.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v52.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, 0
; 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 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v48.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v53.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v50.l, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v52.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v55.l, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v54.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v53.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v49.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v48.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v52, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v39.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v51.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v51.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v52.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v50.h, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v39.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v30.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v55
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v49.h, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v51.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v55
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v48.l, v3.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v27.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v49.l, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v27.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v52, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v52.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v50.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v50.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v52, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v24.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v39.h, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.l, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v24.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.l, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v29.h, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v25.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v29.l, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v28.h, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v52, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v29.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v30.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v52, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v27.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v27.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v26.h, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v52, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v6.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v38.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v25.h, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v24.l, v7.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v23.h, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v52, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v37.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v23.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v23.h, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v22.h, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v36.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v23.l, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v21.h, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v55
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v52, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.h, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v52, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v35.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v21.l, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v21.h, v9.h
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v20.h, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v21.l, v10.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v34.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.h, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v52, v12
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v10.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v20.h, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v52, v13
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v33.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v19.l, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v19.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v19.l, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v18.h, v12.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v55
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.l, 3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v32.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v14
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v18.h, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v15
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v17.l, v13.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v31.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v17.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.h, v14.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v13.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v55
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v14.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v55
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v15.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v55
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v52, v18
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.h, v14.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v52, v17
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v52, v15
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -51295,32 +51088,32 @@ define <64 x i8> @bitcast_v8f64_to_v64i8(<8 x double> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
@@ -51339,26 +51132,26 @@ define <64 x i8> @bitcast_v8f64_to_v64i8(<8 x double> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB84_2: ; %Flow
@@ -51383,26 +51176,26 @@ define <64 x i8> @bitcast_v8f64_to_v64i8(<8 x double> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB84_4: ; %end
@@ -51410,156 +51203,135 @@ define <64 x i8> @bitcast_v8f64_to_v64i8(<8 x double> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l
+; 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_or_b32_e32 v3, v24, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l
; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16
; GFX11-TRUE16-NEXT: s_clause 0x3
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
@@ -54989,15 +54761,15 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:128
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v31, off, s32 offset:124
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:120
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:116
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:116
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:112
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:108
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:108
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:104
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:100
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:100
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:96
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:92
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:92
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:88
-; GFX11-TRUE16-NEXT: scratch_load_b32 v82, off, s32 offset:132
+; GFX11-TRUE16-NEXT: scratch_load_b32 v81, off, s32 offset:132
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:8
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:16
@@ -55011,84 +54783,82 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:80
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:84
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:76
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:68
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:60
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:68
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:60
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:52
; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:44
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:36
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:28
-; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:20
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:36
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:28
+; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:20
; GFX11-TRUE16-NEXT: s_clause 0x1
-; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:12
+; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:12
; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.l, v27.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v25.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v20.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v18.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.h, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v27.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v15.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v49.l, 8, v17.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v25.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v81.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v80.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v27.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(33)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v64.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(31)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(29)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v65.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(27)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v65.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(25)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v66.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v66.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(23)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v66.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v66.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(21)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v67.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(20)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v67.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v67.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(19)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v68.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v68.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v68.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v68.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v69.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v69.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(15)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v70.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(14)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v70.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v70.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(12)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v71.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v82
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v80.l
+; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v81
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB86_3
; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow
@@ -55100,384 +54870,338 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v54.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v51.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v53.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v53.l
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v0.h, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v50.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v51.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v48.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v49.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v1.h, v50.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v39.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v27.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v2.l, v49.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v6, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v3.l, v48.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v24.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v53.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v52.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v52.l
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v0.l, v54.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v55.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v54.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v49.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v64, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v1.h, v53.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v51.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v64, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v2.l, v51.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v48.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v39.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v48.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v64, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v3.l, v50.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v50.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v30.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v64, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v4.l, v39.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v27.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v38.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v4.l, v29.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v4.h, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v38.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v64, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v5.l, v29.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v23.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v37.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v37.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v64, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v6.l, v27.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v36.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v64, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v7.l, v25.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v21.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v64, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v8.l, v23.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v20.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v19.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v33.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v9.l, v22.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v64.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v64, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v10.l, v21.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v31.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v5.l, v28.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v5.h, v26.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v38.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v6.l, v25.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v6.h, v24.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v37.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v7.l, v23.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v37.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v35.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v8.l, v22.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v36.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v22.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v35.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v11, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v9.l, v21.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v9.h, v21.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16
+; 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: $vgpr37_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v12, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v10.l, v20.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v34.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v10.h, v20.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v11.l, v19.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v19.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v33.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v12.l, v18.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v32.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v12.h, v18.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v54
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v13.h, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v13.l, v17.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v54
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v14.l, v16.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_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: $vgpr16_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v17, v54
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b16 v54.h, v15.l, v16.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v64, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v11.l, v20.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v64.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v16.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v64, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v12.l, v19.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v64.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v64, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v13.l, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v64.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v64, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v14.l, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v64.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v64, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v15.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v64.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v54
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v64, v15
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2
; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v54.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v51.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v53.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v53.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v52.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v52.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, 0
; 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 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v48.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v53.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v50.l, 3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v52.l, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v55.l, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v54.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v53.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v49.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v48.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v52, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v39.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v51.l, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v51.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v52.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v50.h, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v39.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v30.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v55
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v49.h, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v51.l, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v55
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v48.l, v3.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v27.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v49.l, v3.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v27.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v3.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v52, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v52.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v50.l, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v50.h, v2.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v52, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v24.h, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v39.h, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.l, v3.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v24.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.l, v4.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v29.h, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v25.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v4.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v29.l, v5.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v28.l, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v28.h, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v26.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v5.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v52, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v29.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v30.h, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v52, v7
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v27.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v27.h, v5.h
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v26.h, v6.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v52, v8
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v6.h
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v38.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v8, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v25.h, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v30.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v6.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v24.l, v7.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v23.h, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v7.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v52, v9
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v37.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v23.l, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v23.h, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v22.h, v8.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v36.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v23.l, v8.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v21.h, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v55
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v9.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v9.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v52, v10
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v8.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.h, v8.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v52, v11
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v35.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v21.l, v9.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v21.h, v9.h
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v11, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v20.h, v10.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v21.l, v10.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v34.h, 3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.h, 3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v13, v55
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.h, v11.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v52, v12
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v10.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v20.h, v10.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v52, v13
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v33.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v19.l, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v19.h, v11.h
; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v11.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v19.l, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v11.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v18.h, v12.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v13, v55
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.l, 3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v12.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v12.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v32.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v14
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v12.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v18.h, v12.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v55
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v15
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v13.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h
; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v15
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v17.l, v13.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v31.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.l, v13.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v17.h, v13.h
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.h, v14.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v13.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v15.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v55
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v14.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v17
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v55
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v55.h, 0x300, v15.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v16, v55
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v52, v18
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.h, v14.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v52, v17
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v52, v15
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -64573,32 +64297,32 @@ define <64 x i8> @bitcast_v32i16_to_v64i8(<32 x i16> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
@@ -64617,26 +64341,26 @@ define <64 x i8> @bitcast_v32i16_to_v64i8(<32 x i16> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB96_2: ; %Flow
@@ -64669,26 +64393,26 @@ define <64 x i8> @bitcast_v32i16_to_v64i8(<32 x i16> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB96_4: ; %end
@@ -64696,156 +64420,135 @@ define <64 x i8> @bitcast_v32i16_to_v64i8(<32 x i16> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l
+; 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_or_b32_e32 v3, v24, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l
; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16
; GFX11-TRUE16-NEXT: s_clause 0x3
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
@@ -76701,32 +76404,32 @@ define <64 x i8> @bitcast_v32f16_to_v64i8(<32 x half> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16
@@ -76745,26 +76448,26 @@ define <64 x i8> @bitcast_v32f16_to_v64i8(<32 x half> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB104_2: ; %Flow
@@ -76797,26 +76500,26 @@ define <64 x i8> @bitcast_v32f16_to_v64i8(<32 x half> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v26, 24, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v27, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v28, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v29, 24, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v30, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v31, 8, v13
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v32, 24, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v33, 8, v12
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v34, 8, v11
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v36, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 24, v8
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v39, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v52, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v25, 8, v1
; GFX11-TRUE16-NEXT: .LBB104_4: ; %end
@@ -76824,156 +76527,135 @@ define <64 x i8> @bitcast_v32f16_to_v64i8(<32 x half> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v64.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h
; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v18.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v55.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v25, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v54.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.l, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v54, 0xffff, v55
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v17.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v54, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v23, v24
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l
+; 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_or_b32_e32 v3, v24, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff, v25
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l
; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l
; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v49.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v23, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v25
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v39.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v22, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v23
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v37.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v9.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v22
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v21, v24
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v35.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v11.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v20, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l
; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l
; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v20, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v32.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v21
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v31.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v18.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v30.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v19, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v20
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v17.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v18, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v26.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16
; GFX11-TRUE16-NEXT: s_clause 0x3
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
@@ -85692,59 +85374,59 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
@@ -85757,307 +85439,302 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 24, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v14
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v67, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v10
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v65, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 8, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v10
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v5
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v2
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v1
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.h, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v4.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v7.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.h, v6.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.h, v7.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v8.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.h, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v10.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.h, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.h, v13.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.h, v14.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.h, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v81.h, v16.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v8.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.h, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v10.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.h, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.h, v13.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v69.h, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.h, v14.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v87.h, v15.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.h, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.h, v16.h
; GFX11-TRUE16-NEXT: .LBB108_2: ; %Flow
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB108_4
; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v1
; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v4
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2
; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_add_f32 v17, 0x40c00000, v17
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_lshlrev_b32 v1, 16, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v11, 16, v11
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v4, 0x40c00000, v4
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v2, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v2
; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v17, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v17
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v2, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v2
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v17, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT: v_add3_u32 v21, v21, v2, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v27, v20, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_add_f32 v1, 0x40c00000, v1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v27.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v17, 0x7fff
; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v26, v20, v22, vcc_lo
; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v1, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v26, v21, v23, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
+; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v26.h
; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v1, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v27, v21, v23, vcc_lo
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v1
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v26
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v27
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v28, v20, v21, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v4, 16, 1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v3
; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v2
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v28.h
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v17, v22, vcc_lo
; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v4, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v4
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v3
; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v2
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v30, v18, v22, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v20, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v29, v18, v22, vcc_lo
; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v6
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v21 :: v_dual_lshlrev_b32 v6, 16, v6
; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v19, 0x7fff
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v21 :: v_dual_lshlrev_b32 v6, 16, v6
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_add_f32 v20, 0x40c00000, v20
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v30, v17, v23, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v21, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v29, v17, v23
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v21
; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v3, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v3
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v21
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v29.h
; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v3, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v30.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v32, v18, v19, vcc_lo
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v5, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v2
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v2
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v31, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v20, v1
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v32.h
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v3, v4, v23 :: v_dual_add_f32 v18, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v17, v29
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v22
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v31.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v4, v23, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v18, 16, 1
+; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v17, v30
; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v6, 16, 1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v19, v3
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
+; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v18, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v6
; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v6, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
+; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v20, 16, 1
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v4
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v4
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v33, v17, v21, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v32, v17, v21, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v31, v19, v22 :: v_dual_and_b32 v20, 0xffff0000, v5
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v20, 0x40c00000, v20 :: v_dual_lshlrev_b32 v5, 16, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v20, 0x7fff
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v3
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v33, v19, v22, vcc_lo
; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v8, 16, v8
-; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v20
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v33.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v5, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v5
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v20, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v20
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v5, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v17, v21, vcc_lo
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v32.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v34, v17, v21 :: v_dual_add_f32 v19, 0x40c00000, v19
; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v8, 16, 1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v36.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v6, v22, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v34.h
; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v19, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v6, v22, vcc_lo
; GFX11-TRUE16-NEXT: v_add3_u32 v6, v17, v8, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v8
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v19
; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v19, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v19
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v35, v6, v17, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v7
-; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v18, v31
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v34, v5, v22, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v18, v33
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v5, v22, vcc_lo
; GFX11-TRUE16-NEXT: v_bfi_b32 v5, 0xffff, v21, v20
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v35.h
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 24, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 24, v6
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v6
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_dual_add_f32 v20, 0x40c00000, v20 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v34
+; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v36
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v5
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v7, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v7
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v8
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v8
; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v7, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v20, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v49, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v23 :: v_dual_add_f32 v10, 0x40c00000, v10
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v38, v19, v21, vcc_lo
; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v20, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v20
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v23 :: v_dual_add_f32 v10, 0x40c00000, v10
; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v17, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v17
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v10, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v17, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v10, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v18, v22, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v10
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v49.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v39, v7, v21, vcc_lo
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v38.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v37, v7, v21, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v12
; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v48, v19, v22 :: v_dual_lshlrev_b32 v7, 16, v9
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v39, v19, v22, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v10
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v37.h
; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_bfe_u32 v22, v21, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v21
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, 0x400000, v21
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v19, v48
+; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v19, v39
; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v12, 16, 1
; GFX11-TRUE16-NEXT: v_add3_u32 v22, v22, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v9
; GFX11-TRUE16-NEXT: v_or_b32_e32 v50, 0x400000, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 24, v10
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v49
; GFX11-TRUE16-NEXT: v_add3_u32 v24, v24, v12, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v54, v22, v37, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v52, v22, v48 :: v_dual_add_f32 v9, 0x40c00000, v23
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v14
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v7 :: v_dual_lshlrev_b32 v14, 16, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v10
+; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v7, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v7
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v14
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14
+; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v9, 0x7fff
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v10
; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v7, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v38, 0xffff0000, v11
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v54.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_lshlrev_b32 v11, 16, v11
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v65, v19, v25, vcc_lo
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v52.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v55, v19, v25, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v9
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v38
-; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v14, 16, 1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v10
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v52, v24, v50 :: v_dual_add_f32 v9, 0x40c00000, v23
; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v21, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v7, v52
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v9, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v14, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v53, v24, v50, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v7, v53
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v20, v23, vcc_lo
; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v11
; GFX11-TRUE16-NEXT: v_add3_u32 v11, v19, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v9, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v21
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v55, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v67, 8, v12
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v20, v23, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v55.h
; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v65.h
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v65, 24, v12
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 8, v12
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v11, v11, v19, vcc_lo
; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_add3_u32 v21, v23, v7, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v7
; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v13
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v19, 16, 1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v23
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v71, v21, v22, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v20, v9
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v23 :: v_dual_cndmask_b32 v70, v21, v22
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_add3_u32 v21, v24, v19, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v19
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT: v_add3_u32 v23, v25, v14, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v14
; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v13, 0x40c00000, v13 :: v_dual_cndmask_b32 v66, v21, v22
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v67, v21, v22, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v7
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_add3_u32 v14, v25, v7, 0x7fff
; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v16
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v68, v23, v24, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v69, v23, v24, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v13, 16, 1
-; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v20, v9
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v9
; GFX11-TRUE16-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v68.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v69.h
; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v7, v14, v19 :: v_dual_add_f32 v14, 0x40c00000, v21
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v15
; GFX11-TRUE16-NEXT: v_add3_u32 v19, v23, v13, 0x7fff
@@ -86067,42 +85744,42 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v16
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v86, v19, v23, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v85, v19, v23, vcc_lo
; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v16, 0x7fff
; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v21, 16, 1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v14, 16, 1
; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v21
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, 0x400000, v21
; GFX11-TRUE16-NEXT: v_add3_u32 v23, v23, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v82, v13, v25, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v83, v13, v25, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT: v_add3_u32 v19, v24, v14, 0x7fff
; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v14
-; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v15, 16, 1
+; GFX11-TRUE16-NEXT: v_bfe_u32 v48, v15, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v15
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v85, v23, v38, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v87, v23, v49, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v86.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v13, v37, v15, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v71.h
-; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v22, v66
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v81, v19, v24, vcc_lo
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v85.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v13, v48, v15, 0x7fff
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v70.h
+; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v22, v67
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v86, v19, v24, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v82.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v85.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v83.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v87.h
; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v23, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 24, v14
; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v13, v13, v16, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfi_b32 v16, 0xffff, v19, v81
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v53, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v9
+; GFX11-TRUE16-NEXT: v_bfi_b32 v16, 0xffff, v19, v86
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 8, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_bfi_b32 v15, 0xffff, v15, v13
; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v21, v7
; GFX11-TRUE16-NEXT: v_bfi_b32 v7, 0xffff, v18, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 24, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v38, 8, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v16
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v16
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
@@ -86111,159 +85788,142 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 8, v15
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v15
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v7
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v7
; GFX11-TRUE16-NEXT: .LBB108_4: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v28.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v113.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v112.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v26.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v2.h, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v103.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, 0
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v1.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v102.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.l, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v101.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v24
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v24.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v27.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v26.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v112.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v103.l
; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v23.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v2.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v8, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v6
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v36.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v3.h, v4.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v23.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v31.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v102.l
+; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v101.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v30.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v99.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v100.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v3.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v8, v24
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v22.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v24, v3
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v5.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v34.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v99.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v6.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v98.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.h, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v4.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v33.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v97.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v22.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v49.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v6.h, v7.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v12
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v33.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v32.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v97.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v6.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v35.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v87.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v7.l, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v65.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v7.h, v8.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v84.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v69.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v24
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v83.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v34.h
-; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v39.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v6.h
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v7.l, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v38.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v96.l
+; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v84.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v36.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v35.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v82.l
+; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v20.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v8.h
+; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v9.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v55.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v81.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v10.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v8.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.h, v10.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v71.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v10.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v48.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v70.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v10.h, v11.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v14, v24
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v67.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v10.l, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v54.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v11.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v64.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v86.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v14, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v16
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v12.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v39.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v37.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v71.l
+; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v19.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v10.h
+; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v11.l, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v70.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v68.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v12.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v66.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v53.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h
; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v52.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v55.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v12.h, v13.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v16, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v19
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v65.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v12.l, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v68.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v53.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v16, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v13.l, v14.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v85.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v13.h, v14.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v51.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v16, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v12.h
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v13.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v85.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v64.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v14.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v54.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v67.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v69.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v51.l
+; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v17.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v14.h
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v15.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v87.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v50.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v14.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v66.h
-; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v14.h, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v82.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v38.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff, v18
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v14.l, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v14.h, v16.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v18, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v19
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v15.l, v15.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v16.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v16.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v49.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14
+; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v86.h
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v16.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h
+; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v83.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v17, v24
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v17, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15
+; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v16.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v17.l, v17.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16
; GFX11-TRUE16-NEXT: s_clause 0x3
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll
index 6fe66655de3d6..4c485768bcbbf 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll
@@ -3065,13 +3065,12 @@ define i64 @bitcast_v8i8_to_i64(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8
@@ -3085,61 +3084,53 @@ define i64 @bitcast_v8i8_to_i64(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2
; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -6214,13 +6205,12 @@ define double @bitcast_v8i8_to_f64(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8
@@ -6234,61 +6224,53 @@ define double @bitcast_v8i8_to_f64(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2
; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -9063,13 +9045,12 @@ define <2 x i32> @bitcast_v8i8_to_v2i32(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8
@@ -9083,61 +9064,53 @@ define <2 x i32> @bitcast_v8i8_to_v2i32(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2
; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -11603,13 +11576,12 @@ define <2 x float> @bitcast_v8i8_to_v2f32(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8
@@ -11623,61 +11595,53 @@ define <2 x float> @bitcast_v8i8_to_v2f32(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2
; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -13829,13 +13793,12 @@ define <4 x i16> @bitcast_v8i8_to_v4i16(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8
@@ -13849,61 +13812,53 @@ define <4 x i16> @bitcast_v8i8_to_v4i16(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB98_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_2
; GFX11-TRUE16-NEXT: .LBB98_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -15655,13 +15610,12 @@ define <4 x half> @bitcast_v8i8_to_v4f16(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8
@@ -15675,61 +15629,53 @@ define <4 x half> @bitcast_v8i8_to_v4f16(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB106_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB106_2
; GFX11-TRUE16-NEXT: .LBB106_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -16966,13 +16912,12 @@ define <4 x bfloat> @bitcast_v8i8_to_v4bf16(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v7.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v8
@@ -16986,61 +16931,53 @@ define <4 x bfloat> @bitcast_v8i8_to_v4bf16(<8 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT: .LBB110_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v3.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v4.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.h
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v2.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v4
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB110_2
; GFX11-TRUE16-NEXT: .LBB110_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v3.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v3.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.h, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v2, v5
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll
index e5245f7bd71d3..879e8520d8e18 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll
@@ -1102,15 +1102,16 @@ define <3 x i32> @bitcast_v12i8_to_v3i32(<12 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v8.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v8.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v11.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12
@@ -1125,80 +1126,74 @@ define <3 x i32> @bitcast_v12i8_to_v3i32(<12 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB6_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v0.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v7.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v4.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v3.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v1.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v7.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v3.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v2.l, v3.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v2.l, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v7.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v2
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB6_2
; GFX11-TRUE16-NEXT: .LBB6_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v7.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v7.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v5.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.h, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v5.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.l, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v7.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v4.l, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v7.h
; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.h, v2.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -4241,15 +4236,16 @@ define <3 x float> @bitcast_v12i8_to_v3f32(<12 x i8> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v8.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v8.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v11.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12
@@ -4264,80 +4260,74 @@ define <3 x float> @bitcast_v12i8_to_v3f32(<12 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB22_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v7.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v5.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.h
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v0.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v5.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v7.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v4.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v3.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v1.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v7.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v3.h
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v2.l, v3.l
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v2.l, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v7.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v7
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v2
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB22_2
; GFX11-TRUE16-NEXT: .LBB22_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v7.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v7.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v5.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.h, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v5.h, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.l, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v7.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v4.l, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v5
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v7.h
; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v7
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v7
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.h, v2.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v7.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -6885,16 +6875,16 @@ define <6 x bfloat> @bitcast_v12i8_to_v6bf16(<12 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v9.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v10.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v10.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12
@@ -6909,80 +6899,74 @@ define <6 x bfloat> @bitcast_v12i8_to_v6bf16(<12 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB36_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v9.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v6.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v1.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v2.l, v4.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v4.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v2.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v7
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB36_2
; GFX11-TRUE16-NEXT: .LBB36_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v9.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v8.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.l, v0.l
; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.h, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.h
; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.h, v2.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -8651,16 +8635,16 @@ define <6 x half> @bitcast_v12i8_to_v6f16(<12 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v9.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v10.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v10.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12
@@ -8675,80 +8659,74 @@ define <6 x half> @bitcast_v12i8_to_v6f16(<12 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB40_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v9.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v6.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v1.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v2.l, v4.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v4.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v2.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v7
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB40_2
; GFX11-TRUE16-NEXT: .LBB40_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v9.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v8.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.l, v0.l
; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.h, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.h
; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.h, v2.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -10065,16 +10043,16 @@ define <6 x i16> @bitcast_v12i8_to_v6i16(<12 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v9.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v10.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v11.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v10.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l
; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12
@@ -10089,80 +10067,74 @@ define <6 x i16> @bitcast_v12i8_to_v6i16(<12 x i8> %a, i32 %b) {
; GFX11-TRUE16-NEXT: .LBB44_3: ; %cmp.false
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v9.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v8.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.l
-; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v5.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v10.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v6.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v6.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v1.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v1.l, v4.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v5
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v6, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v2.l, v4.l
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v0
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v4.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v5
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v2.l, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v7
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB44_2
; GFX11-TRUE16-NEXT: .LBB44_4: ; %cmp.true
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v9.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v8.h, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.h, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v8.l, 3
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v6.l, 3
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3
; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.l, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.l, v0.l
; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.h, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, 0
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.h
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v0.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v6
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.h
; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v1.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v9
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v2.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v9
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.h, v2.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v1
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v2.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-llvm-debuginfo-analyzer.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-llvm-debuginfo-analyzer.ll
index 89fc6c062c29d..d6922bc09ff0a 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-llvm-debuginfo-analyzer.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-llvm-debuginfo-analyzer.ll
@@ -1,3 +1,4 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc %s -o %t.o -mcpu=gfx1030 -filetype=obj -O0
; RUN: llvm-debuginfo-analyzer %t.o --print=all --attribute=all | FileCheck %s
diff --git a/llvm/test/CodeGen/AMDGPU/atomic_optimizations_global_pointer.ll b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_global_pointer.ll
index 4cc39d93854a0..1d3368b036d0d 100644
--- a/llvm/test/CodeGen/AMDGPU/atomic_optimizations_global_pointer.ll
+++ b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_global_pointer.ll
@@ -9022,13 +9022,12 @@ define amdgpu_kernel void @uniform_or_i16(ptr addrspace(1) %result, ptr addrspac
; GFX1164-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s7, v0
; GFX1164-TRUE16-NEXT: .LBB15_2:
; GFX1164-TRUE16-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX1164-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX1164-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1164-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1164-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX1164-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
+; GFX1164-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX1164-TRUE16-NEXT: v_readfirstlane_b32 s2, v0
; GFX1164-TRUE16-NEXT: v_cndmask_b16 v0.l, s6, 0, vcc
-; GFX1164-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1164-TRUE16-NEXT: v_or_b16 v0.l, s2, v0.l
; GFX1164-TRUE16-NEXT: s_mov_b32 s2, -1
; GFX1164-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0
@@ -9101,13 +9100,12 @@ define amdgpu_kernel void @uniform_or_i16(ptr addrspace(1) %result, ptr addrspac
; GFX1132-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s6, v0
; GFX1132-TRUE16-NEXT: .LBB15_2:
; GFX1132-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX1132-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX1132-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1132-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1132-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX1132-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
+; GFX1132-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX1132-TRUE16-NEXT: v_readfirstlane_b32 s2, v0
; GFX1132-TRUE16-NEXT: v_cndmask_b16 v0.l, s4, 0, vcc_lo
-; GFX1132-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-TRUE16-NEXT: v_or_b16 v0.l, s2, v0.l
; GFX1132-TRUE16-NEXT: s_mov_b32 s2, -1
; GFX1132-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0
@@ -9180,13 +9178,12 @@ define amdgpu_kernel void @uniform_or_i16(ptr addrspace(1) %result, ptr addrspac
; GFX1264-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s7, v0
; GFX1264-TRUE16-NEXT: .LBB15_2:
; GFX1264-TRUE16-NEXT: s_or_b64 exec, exec, s[4:5]
-; GFX1264-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX1264-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1264-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1264-TRUE16-NEXT: s_wait_kmcnt 0x0
; GFX1264-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
+; GFX1264-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX1264-TRUE16-NEXT: v_readfirstlane_b32 s2, v0
; GFX1264-TRUE16-NEXT: v_cndmask_b16 v0.l, s6, 0, vcc
-; GFX1264-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1264-TRUE16-NEXT: v_or_b16 v0.l, s2, v0.l
; GFX1264-TRUE16-NEXT: s_mov_b32 s2, -1
; GFX1264-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], null
@@ -9259,13 +9256,12 @@ define amdgpu_kernel void @uniform_or_i16(ptr addrspace(1) %result, ptr addrspac
; GFX1232-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s6, v0
; GFX1232-TRUE16-NEXT: .LBB15_2:
; GFX1232-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s5
-; GFX1232-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX1232-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1232-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1232-TRUE16-NEXT: s_wait_kmcnt 0x0
; GFX1232-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
+; GFX1232-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX1232-TRUE16-NEXT: v_readfirstlane_b32 s2, v0
; GFX1232-TRUE16-NEXT: v_cndmask_b16 v0.l, s4, 0, vcc_lo
-; GFX1232-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1232-TRUE16-NEXT: v_or_b16 v0.l, s2, v0.l
; GFX1232-TRUE16-NEXT: s_mov_b32 s2, -1
; GFX1232-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], null
@@ -9662,12 +9658,11 @@ define amdgpu_kernel void @uniform_add_i16(ptr addrspace(1) %result, ptr addrspa
; GFX1164-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s11, v2
; GFX1164-TRUE16-NEXT: .LBB16_4: ; %Flow
; GFX1164-TRUE16-NEXT: s_or_b64 exec, exec, s[8:9]
-; GFX1164-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX1164-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1164-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1164-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX1164-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
+; GFX1164-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1164-TRUE16-NEXT: v_readfirstlane_b32 s2, v0
-; GFX1164-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1164-TRUE16-NEXT: v_mad_u16 v0.l, s10, v4.l, s2
; GFX1164-TRUE16-NEXT: s_mov_b32 s2, -1
; GFX1164-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0
@@ -9789,12 +9784,11 @@ define amdgpu_kernel void @uniform_add_i16(ptr addrspace(1) %result, ptr addrspa
; GFX1132-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s2, v2
; GFX1132-TRUE16-NEXT: .LBB16_4: ; %Flow
; GFX1132-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s9
-; GFX1132-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX1132-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1132-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1132-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX1132-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
+; GFX1132-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1132-TRUE16-NEXT: v_readfirstlane_b32 s2, v0
-; GFX1132-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-TRUE16-NEXT: v_mad_u16 v0.l, s8, v4.l, s2
; GFX1132-TRUE16-NEXT: s_mov_b32 s2, -1
; GFX1132-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], 0
@@ -9916,13 +9910,12 @@ define amdgpu_kernel void @uniform_add_i16(ptr addrspace(1) %result, ptr addrspa
; GFX1264-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s11, v2
; GFX1264-TRUE16-NEXT: .LBB16_4: ; %Flow
; GFX1264-TRUE16-NEXT: s_or_b64 exec, exec, s[8:9]
-; GFX1264-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX1264-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1264-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1264-TRUE16-NEXT: s_wait_kmcnt 0x0
; GFX1264-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
+; GFX1264-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX1264-TRUE16-NEXT: v_readfirstlane_b32 s2, v0
; GFX1264-TRUE16-NEXT: s_wait_alu 0xf1ff
-; GFX1264-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1264-TRUE16-NEXT: v_mad_u16 v0.l, s10, v4.l, s2
; GFX1264-TRUE16-NEXT: s_mov_b32 s2, -1
; GFX1264-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], null
@@ -10048,13 +10041,12 @@ define amdgpu_kernel void @uniform_add_i16(ptr addrspace(1) %result, ptr addrspa
; GFX1232-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s2, v2
; GFX1232-TRUE16-NEXT: .LBB16_4: ; %Flow
; GFX1232-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s9
-; GFX1232-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX1232-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1232-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1232-TRUE16-NEXT: s_wait_kmcnt 0x0
; GFX1232-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
+; GFX1232-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX1232-TRUE16-NEXT: v_readfirstlane_b32 s2, v0
; GFX1232-TRUE16-NEXT: s_wait_alu 0xf1ff
-; GFX1232-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1232-TRUE16-NEXT: v_mad_u16 v0.l, s8, v4.l, s2
; GFX1232-TRUE16-NEXT: s_mov_b32 s2, -1
; GFX1232-TRUE16-NEXT: buffer_store_b16 v0, off, s[0:3], null
@@ -10734,15 +10726,15 @@ define amdgpu_kernel void @uniform_fadd_f16(ptr addrspace(1) %result, ptr addrsp
; GFX1164-TRUE16-NEXT: s_mov_b64 s[2:3], 0
; GFX1164-TRUE16-NEXT: .LBB18_1: ; %atomicrmw.start
; GFX1164-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1164-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1164-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX1164-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s9, v1
+; GFX1164-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1164-TRUE16-NEXT: v_add_f16_e32 v0.l, s8, v0.l
; GFX1164-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1164-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX1164-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s9, v0
-; GFX1164-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX1164-TRUE16-NEXT: v_and_or_b32 v0, v1, s10, v0
; GFX1164-TRUE16-NEXT: v_mov_b32_e32 v3, v1
+; GFX1164-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX1164-TRUE16-NEXT: v_mov_b32_e32 v2, v0
; GFX1164-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[2:3], off, s[4:7], 0 glc
; GFX1164-TRUE16-NEXT: s_waitcnt vmcnt(0)
@@ -10828,14 +10820,14 @@ define amdgpu_kernel void @uniform_fadd_f16(ptr addrspace(1) %result, ptr addrsp
; GFX1132-TRUE16-NEXT: s_mov_b32 s6, -1
; GFX1132-TRUE16-NEXT: .LBB18_1: ; %atomicrmw.start
; GFX1132-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1132-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1132-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX1132-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s2, v1
+; GFX1132-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1132-TRUE16-NEXT: v_add_f16_e32 v0.l, s8, v0.l
; GFX1132-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1132-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX1132-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s2, v0
-; GFX1132-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1132-TRUE16-NEXT: v_and_or_b32 v0, v1, s3, v0
+; GFX1132-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1132-TRUE16-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
; GFX1132-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[2:3], off, s[4:7], 0 glc
; GFX1132-TRUE16-NEXT: s_waitcnt vmcnt(0)
@@ -10920,15 +10912,15 @@ define amdgpu_kernel void @uniform_fadd_f16(ptr addrspace(1) %result, ptr addrsp
; GFX1264-TRUE16-NEXT: s_mov_b64 s[2:3], 0
; GFX1264-TRUE16-NEXT: .LBB18_1: ; %atomicrmw.start
; GFX1264-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1264-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1264-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX1264-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s9, v1
+; GFX1264-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1264-TRUE16-NEXT: v_add_f16_e32 v0.l, s8, v0.l
; GFX1264-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1264-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX1264-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s9, v0
-; GFX1264-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX1264-TRUE16-NEXT: v_and_or_b32 v0, v1, s10, v0
; GFX1264-TRUE16-NEXT: v_mov_b32_e32 v3, v1
+; GFX1264-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX1264-TRUE16-NEXT: v_mov_b32_e32 v2, v0
; GFX1264-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[2:3], off, s[4:7], null th:TH_ATOMIC_RETURN scope:SCOPE_SYS
; GFX1264-TRUE16-NEXT: s_wait_loadcnt 0x0
@@ -11014,14 +11006,14 @@ define amdgpu_kernel void @uniform_fadd_f16(ptr addrspace(1) %result, ptr addrsp
; GFX1232-TRUE16-NEXT: s_mov_b32 s6, -1
; GFX1232-TRUE16-NEXT: .LBB18_1: ; %atomicrmw.start
; GFX1232-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
-; GFX1232-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1232-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX1232-TRUE16-NEXT: v_lshrrev_b32_e32 v0, s2, v1
+; GFX1232-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1232-TRUE16-NEXT: v_add_f16_e32 v0.l, s8, v0.l
; GFX1232-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX1232-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX1232-TRUE16-NEXT: v_lshlrev_b32_e32 v0, s2, v0
-; GFX1232-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1232-TRUE16-NEXT: v_and_or_b32 v0, v1, s3, v0
+; GFX1232-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1232-TRUE16-NEXT: v_dual_mov_b32 v3, v1 :: v_dual_mov_b32 v2, v0
; GFX1232-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[2:3], off, s[4:7], null th:TH_ATOMIC_RETURN scope:SCOPE_SYS
; GFX1232-TRUE16-NEXT: s_wait_loadcnt 0x0
diff --git a/llvm/test/CodeGen/AMDGPU/bf16.ll b/llvm/test/CodeGen/AMDGPU/bf16.ll
index 505ddc8c3b575..10e523d1a0cf1 100644
--- a/llvm/test/CodeGen/AMDGPU/bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/bf16.ll
@@ -37774,9 +37774,10 @@ define bfloat @v_uitofp_i16_to_bf16(i16 %x) {
; GFX11TRUE16-LABEL: v_uitofp_i16_to_bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11TRUE16-NEXT: v_cvt_f32_u32_e32 v0, v0
+; GFX11TRUE16-NEXT: v_cvt_f32_u32_e32 v0, v1
; GFX11TRUE16-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
@@ -40750,12 +40751,11 @@ define amdgpu_ps i32 @s_select_bf16(bfloat inreg %a, bfloat inreg %b, i32 %c) {
;
; GFX11TRUE16-LABEL: s_select_bf16:
; GFX11TRUE16: ; %bb.0:
+; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, s0
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
-; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.l, s0
-; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11TRUE16-NEXT: v_cndmask_b16 v0.l, s1, v0.l, vcc_lo
-; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
+; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11TRUE16-NEXT: v_cndmask_b16 v0.l, s1, v1.l, vcc_lo
; GFX11TRUE16-NEXT: v_readfirstlane_b32 s0, v0
; GFX11TRUE16-NEXT: ; return to shader part epilog
;
diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll
index f4b432dce8c8a..0ceb9019eb990 100644
--- a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll
+++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fadd.ll
@@ -3443,15 +3443,14 @@ define half @buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__amdgpu_no_fine_gr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v1.l, v1.l, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1
; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], null offen th:TH_ATOMIC_RETURN
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
@@ -3569,14 +3568,13 @@ define half @buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__amdgpu_no_fine_gr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v1.l, v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1
; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], 0 offen glc
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
@@ -3884,15 +3882,14 @@ define void @buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset__amdgpu_no_fine_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v1.l, v1.l, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1
; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
@@ -4007,14 +4004,13 @@ define void @buffer_fat_ptr_agent_atomic_fadd_noret_f16__offset__amdgpu_no_fine_
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v1.l, v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1
; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
@@ -4328,15 +4324,14 @@ define half @buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall__amdgpu
; GFX12-TRUE16-NEXT: ; Child Loop BB15_4 Depth 2
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v6, v4, v7
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v6.h, 0
; GFX12-TRUE16-NEXT: s_mov_b32 s2, exec_lo
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v6.l, v6.l, v5.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v6, v4, v6
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v6, v7, v11, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
; GFX12-TRUE16-NEXT: .LBB15_4: ; Parent Loop BB15_3 Depth=1
; GFX12-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2
@@ -4556,15 +4551,14 @@ define half @buffer_fat_ptr_agent_atomic_fadd_ret_f16__offset__waterfall__amdgpu
; GFX11-TRUE16-NEXT: ; Child Loop BB15_4 Depth 2
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v6, v4, v7
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, 0
; GFX11-TRUE16-NEXT: s_mov_b32 s2, exec_lo
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v6.l, v6.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, v4, v6
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v6, v7, v11, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v9, v7 :: v_dual_mov_b32 v8, v6
; GFX11-TRUE16-NEXT: .LBB15_4: ; Parent Loop BB15_3 Depth=1
; GFX11-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2
diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll
index 6f1675edbe58a..cad4c39eaf39f 100644
--- a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll
+++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmax.ll
@@ -2512,16 +2512,16 @@ define half @buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__amdgpu_no_fine_gr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v1.l, v1.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v1.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1
; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], null offen th:TH_ATOMIC_RETURN
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
@@ -2640,20 +2640,19 @@ define half @buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__amdgpu_no_fine_gr
; GFX11-TRUE16-NEXT: buffer_load_b32 v2, v5, s[0:3], 0 offen
; GFX11-TRUE16-NEXT: s_not_b32 s6, s5
; GFX11-TRUE16-NEXT: s_mov_b32 s5, 0
-; GFX11-TRUE16-NEXT: .p2align 6
; GFX11-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v1.l, v1.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v1.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1
; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], 0 offen glc
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
@@ -2973,16 +2972,16 @@ define void @buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset__amdgpu_no_fine_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v1.l, v1.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v1.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1
; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
@@ -3098,20 +3097,19 @@ define void @buffer_fat_ptr_agent_atomic_fmax_noret_f16__offset__amdgpu_no_fine_
; GFX11-TRUE16-NEXT: buffer_load_b32 v2, v3, s[0:3], 0 offen
; GFX11-TRUE16-NEXT: s_not_b32 s6, s5
; GFX11-TRUE16-NEXT: s_mov_b32 s5, 0
-; GFX11-TRUE16-NEXT: .p2align 6
; GFX11-TRUE16-NEXT: .LBB11_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v1.l, v1.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v1.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1
; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
@@ -3437,16 +3435,16 @@ define half @buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall__amdgpu
; GFX12-TRUE16-NEXT: ; Child Loop BB12_4 Depth 2
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v9, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: s_mov_b32 s2, exec_lo
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v4.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v4.h, v4.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v9, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v11, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_dual_mov_b32 v8, v6 :: v_dual_mov_b32 v7, v5
; GFX12-TRUE16-NEXT: .LBB12_4: ; Parent Loop BB12_3 Depth=1
; GFX12-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2
@@ -3672,16 +3670,16 @@ define half @buffer_fat_ptr_agent_atomic_fmax_ret_f16__offset__waterfall__amdgpu
; GFX11-TRUE16-NEXT: ; Child Loop BB12_4 Depth 2
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v9, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: s_mov_b32 s2, exec_lo
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v4.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v4.h, v4.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v9, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v11, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, v6 :: v_dual_mov_b32 v7, v5
; GFX11-TRUE16-NEXT: .LBB12_4: ; Parent Loop BB12_3 Depth=1
; GFX11-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2
diff --git a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll
index acb27be1846b9..6275afd2c6994 100644
--- a/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll
+++ b/llvm/test/CodeGen/AMDGPU/buffer-fat-pointer-atomicrmw-fmin.ll
@@ -2512,16 +2512,16 @@ define half @buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__amdgpu_no_fine_gr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v1.l, v1.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v1.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1
; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], null offen th:TH_ATOMIC_RETURN
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
@@ -2640,20 +2640,19 @@ define half @buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__amdgpu_no_fine_gr
; GFX11-TRUE16-NEXT: buffer_load_b32 v2, v5, s[0:3], 0 offen
; GFX11-TRUE16-NEXT: s_not_b32 s6, s5
; GFX11-TRUE16-NEXT: s_mov_b32 s5, 0
-; GFX11-TRUE16-NEXT: .p2align 6
; GFX11-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v1.l, v1.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v1.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1
; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[3:4], v5, s[0:3], 0 offen glc
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
@@ -2973,16 +2972,16 @@ define void @buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset__amdgpu_no_fine_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v1.l, v1.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v1.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1
; GFX12-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], null offen th:TH_ATOMIC_RETURN
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
@@ -3098,20 +3097,19 @@ define void @buffer_fat_ptr_agent_atomic_fmin_noret_f16__offset__amdgpu_no_fine_
; GFX11-TRUE16-NEXT: buffer_load_b32 v2, v3, s[0:3], 0 offen
; GFX11-TRUE16-NEXT: s_not_b32 s6, s5
; GFX11-TRUE16-NEXT: s_mov_b32 s5, 0
-; GFX11-TRUE16-NEXT: .p2align 6
; GFX11-TRUE16-NEXT: .LBB11_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v1, s4, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v1.l, v1.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v1.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, s4, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v1, v2, s6, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1
; GFX11-TRUE16-NEXT: buffer_atomic_cmpswap_b32 v[4:5], v3, s[0:3], 0 offen glc
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
@@ -3437,16 +3435,16 @@ define half @buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall__amdgpu
; GFX12-TRUE16-NEXT: ; Child Loop BB12_4 Depth 2
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v9, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: s_mov_b32 s2, exec_lo
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v4.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v4.h, v4.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v9, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v11, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_dual_mov_b32 v8, v6 :: v_dual_mov_b32 v7, v5
; GFX12-TRUE16-NEXT: .LBB12_4: ; Parent Loop BB12_3 Depth=1
; GFX12-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2
@@ -3672,16 +3670,16 @@ define half @buffer_fat_ptr_agent_atomic_fmin_ret_f16__offset__waterfall__amdgpu
; GFX11-TRUE16-NEXT: ; Child Loop BB12_4 Depth 2
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v9, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: s_mov_b32 s2, exec_lo
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v4.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v4.h, v4.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v9, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v11, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, v6 :: v_dual_mov_b32 v7, v5
; GFX11-TRUE16-NEXT: .LBB12_4: ; Parent Loop BB12_3 Depth=1
; GFX11-TRUE16-NEXT: ; => This Inner Loop Header: Depth=2
diff --git a/llvm/test/CodeGen/AMDGPU/calling-conventions.ll b/llvm/test/CodeGen/AMDGPU/calling-conventions.ll
index ff80250bfc880..2db7b28c7de97 100644
--- a/llvm/test/CodeGen/AMDGPU/calling-conventions.ll
+++ b/llvm/test/CodeGen/AMDGPU/calling-conventions.ll
@@ -2745,6 +2745,15 @@ define amdgpu_cs void @amdgpu_cs_v32i1(<32 x i1> %arg0) {
;
; GFX11-TRUE16-LABEL: amdgpu_cs_v32i1:
; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: v_and_b16 v26.l, v26.l, 1
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 1, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v24.l, v24.l, 1
+; GFX11-TRUE16-NEXT: v_and_b16 v20.h, v22.l, 1
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 1, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v20.l, v20.l, 1
+; GFX11-TRUE16-NEXT: v_and_b16 v17.h, v18.l, 1
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 1, v17.l
+; GFX11-TRUE16-NEXT: v_and_b16 v16.l, v16.l, 1
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, v10.l, 1
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 1, v9.l
; GFX11-TRUE16-NEXT: v_and_b16 v8.l, v8.l, 1
@@ -2754,6 +2763,18 @@ define amdgpu_cs void @amdgpu_cs_v32i1(<32 x i1> %arg0) {
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, v2.l, 1
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 1, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, v0.l, 1
+; GFX11-TRUE16-NEXT: v_and_b16 v30.l, v30.l, 1
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 1, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v28.l, v28.l, 1
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 3, v27.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.l, 2, v26.l
+; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v24.l, v25.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 3, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 2, v20.h
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v20.l, v21.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 3, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 2, v17.h
+; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v17.l
; GFX11-TRUE16-NEXT: v_and_b16 v14.l, v14.l, 1
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 1, v13.l
; GFX11-TRUE16-NEXT: v_and_b16 v12.l, v12.l, 1
@@ -2766,15 +2787,15 @@ define amdgpu_cs void @amdgpu_cs_v32i1(<32 x i1> %arg0) {
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 3, v3.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 2, v1.h
; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v25.h, v26.l, 1
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 1, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, v24.l, 1
-; GFX11-TRUE16-NEXT: v_and_b16 v22.l, v22.l, 1
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 1, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v20.l, v20.l, 1
-; GFX11-TRUE16-NEXT: v_and_b16 v18.l, v18.l, 1
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 1, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, v16.l, 1
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 3, v31.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 2, v30.l
+; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v29.l
+; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v26.l
+; GFX11-TRUE16-NEXT: v_and_b16 v21.h, v22.h, 3
+; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v22.l, v20.h
+; GFX11-TRUE16-NEXT: v_and_b16 v16.h, v16.h, 3
+; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v18.h, v17.h
+; GFX11-TRUE16-NEXT: v_and_b16 v16.l, v16.l, 3
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 3, v15.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 2, v14.l
; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v12.l, v13.l
@@ -2784,65 +2805,42 @@ define amdgpu_cs void @amdgpu_cs_v32i1(<32 x i1> %arg0) {
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, v0.h, 3
; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.h, v1.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, v0.l, 3
-; GFX11-TRUE16-NEXT: v_and_b16 v30.l, v30.l, 1
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 1, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v28.l, v28.l, 1
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 3, v27.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 2, v25.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v25.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 3, v23.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 2, v22.l
-; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v20.l, v21.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 3, v19.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 2, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.l, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v28.h, v29.h
+; GFX11-TRUE16-NEXT: v_and_b16 v24.h, v28.l, 3
+; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v21.h, v25.h
+; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.l, v17.h
; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v12.h, v10.h
; GFX11-TRUE16-NEXT: v_and_b16 v3.l, v8.h, 3
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v6.h
; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v2.l
; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 3, v31.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 2, v30.l
-; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v28.l, v29.l
-; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v25.h
-; GFX11-TRUE16-NEXT: v_and_b16 v24.l, v24.l, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v23.l, v22.l
-; GFX11-TRUE16-NEXT: v_and_b16 v14.h, v16.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v17.h, v18.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, v15.h, 3
+; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v26.h
+; GFX11-TRUE16-NEXT: v_and_b16 v14.h, v19.l, 15
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 4, v16.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, v15.h, 15
; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v2.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.l, v1.l, 15
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 4, v0.h
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, v0.l, 15
-; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v30.h, v28.h
-; GFX11-TRUE16-NEXT: v_and_b16 v24.h, v24.h, 3
-; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v24.l, v22.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v14.h, v18.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 12, v24.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v14.h
; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v16.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 12, v2.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v24.h, v28.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, v20.h, 15
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 4, v2.h
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, v1.h, 15
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v17.l, v2.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 12, v23.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.h, v0.h
; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.l, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2
; GFX11-TRUE16-NEXT: global_store_b32 v[0:1], v0, off
; GFX11-TRUE16-NEXT: s_endpgm
;
diff --git a/llvm/test/CodeGen/AMDGPU/clamp-modifier.ll b/llvm/test/CodeGen/AMDGPU/clamp-modifier.ll
index b9caf8e80bcdf..ccdc0b1bf43c4 100644
--- a/llvm/test/CodeGen/AMDGPU/clamp-modifier.ll
+++ b/llvm/test/CodeGen/AMDGPU/clamp-modifier.ll
@@ -1561,10 +1561,10 @@ define amdgpu_kernel void @v_no_clamp_add_src_v2f16_f16_src(ptr addrspace(1) %ou
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 2, v1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: global_load_d16_b16 v0, v0, s[2:3]
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v0.l, 1.0, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_pk_max_f16 v0, v0, v0 clamp
; GFX11-TRUE16-NEXT: global_store_b32 v1, v0, s[0:1]
; GFX11-TRUE16-NEXT: s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll b/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll
index b5bc09a1684ee..26f204f29f5a4 100644
--- a/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll
+++ b/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll
@@ -946,9 +946,9 @@ define double @v_uitofp_i8_to_f64(i8 %arg0) nounwind {
; GFX11-TRUE16-LABEL: v_uitofp_i8_to_f64:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_cvt_f64_u32_e32 v[0:1], v0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -1770,40 +1770,38 @@ define amdgpu_kernel void @load_v4i8_to_v4f32_2_uses(ptr addrspace(1) noalias %o
; GFX11-TRUE16-LABEL: load_v4i8_to_v4f32_2_uses:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x34
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v5, 0 :: v_dual_and_b32 v0, 0x3ff, v0
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0x3ff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: global_load_b32 v4, v0, s[0:1]
; GFX11-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v4.l, 9
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff00, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff00, v4.h
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 9
+; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff00, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff00, v4.h
; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte3_e32 v3, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 9
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x900, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.h
+; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte2_e32 v2, v4
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h
; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte1_e32 v1, v4
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x900, v0.l
-; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte2_e32 v2, v4
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x900, v0.l
+; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x900, v0.h
; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte0_e32 v0, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v7, v6
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v5, v7
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: s_clause 0x1
-; GFX11-TRUE16-NEXT: global_store_b128 v5, v[0:3], s[0:1]
-; GFX11-TRUE16-NEXT: global_store_b32 v5, v4, s[2:3]
+; GFX11-TRUE16-NEXT: global_store_b128 v6, v[0:3], s[0:1]
+; GFX11-TRUE16-NEXT: global_store_b32 v6, v4, s[2:3]
; GFX11-TRUE16-NEXT: s_endpgm
;
; GFX11-FAKE16-LABEL: load_v4i8_to_v4f32_2_uses:
diff --git a/llvm/test/CodeGen/AMDGPU/dynamic_stackalloc.ll b/llvm/test/CodeGen/AMDGPU/dynamic_stackalloc.ll
index b0439b1f7968f..c5db7a33f70e0 100644
--- a/llvm/test/CodeGen/AMDGPU/dynamic_stackalloc.ll
+++ b/llvm/test/CodeGen/AMDGPU/dynamic_stackalloc.ll
@@ -2536,12 +2536,13 @@ define void @test_dynamic_stackalloc_device_divergent_non_standard_size_i16(i16
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_device_divergent_non_standard_size_i16:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-SDAG-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-SDAG-NEXT: v_mov_b16_e32 v1.l, v0.l
; GFX11-SDAG-NEXT: s_mov_b32 s4, s33
; GFX11-SDAG-NEXT: s_mov_b32 s1, exec_lo
; GFX11-SDAG-NEXT: s_mov_b32 s0, 0
; GFX11-SDAG-NEXT: s_mov_b32 s33, s32
-; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
+; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v1, 2, 15
; GFX11-SDAG-NEXT: s_add_i32 s32, s32, 16
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x7fff0, v0
diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll
index 8c7d5cffe39d9..22dd66118837f 100644
--- a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fadd.ll
@@ -8410,13 +8410,12 @@ define half @flat_agent_atomic_fadd_ret_f16__amdgpu_no_fine_grained_memory(ptr %
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8529,13 +8528,12 @@ define half @flat_agent_atomic_fadd_ret_f16__amdgpu_no_fine_grained_memory(ptr %
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
@@ -8785,13 +8783,12 @@ define half @flat_agent_atomic_fadd_ret_f16__offset12b_pos__amdgpu_no_fine_grain
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8908,13 +8905,12 @@ define half @flat_agent_atomic_fadd_ret_f16__offset12b_pos__amdgpu_no_fine_grain
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
@@ -9171,13 +9167,12 @@ define half @flat_agent_atomic_fadd_ret_f16__offset12b_neg__amdgpu_no_fine_grain
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -9295,13 +9290,12 @@ define half @flat_agent_atomic_fadd_ret_f16__offset12b_neg__amdgpu_no_fine_grain
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
@@ -9557,11 +9551,11 @@ define void @flat_agent_atomic_fadd_noret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -9671,11 +9665,11 @@ define void @flat_agent_atomic_fadd_noret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
@@ -9917,11 +9911,11 @@ define void @flat_agent_atomic_fadd_noret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -10035,11 +10029,11 @@ define void @flat_agent_atomic_fadd_noret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
@@ -10288,11 +10282,11 @@ define void @flat_agent_atomic_fadd_noret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -10407,11 +10401,11 @@ define void @flat_agent_atomic_fadd_noret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
@@ -10651,8 +10645,8 @@ define void @flat_agent_atomic_fadd_noret_f16__offset12b__align4_pos__amdgpu_no_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v4.l, v2.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -10735,8 +10729,8 @@ define void @flat_agent_atomic_fadd_noret_f16__offset12b__align4_pos__amdgpu_no_
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v4.l, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
@@ -10925,10 +10919,9 @@ define half @flat_agent_atomic_fadd_ret_f16__offset12b_pos__align4__amdgpu_no_fi
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v4.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -11014,10 +11007,9 @@ define half @flat_agent_atomic_fadd_ret_f16__offset12b_pos__align4__amdgpu_no_fi
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
@@ -11220,13 +11212,12 @@ define half @flat_system_atomic_fadd_ret_f16__offset12b_pos__amdgpu_no_fine_grai
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -11345,13 +11336,12 @@ define half @flat_system_atomic_fadd_ret_f16__offset12b_pos__amdgpu_no_fine_grai
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
@@ -11610,11 +11600,11 @@ define void @flat_system_atomic_fadd_noret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -11730,11 +11720,11 @@ define void @flat_system_atomic_fadd_noret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll
index 56ad91dd59ffb..1dc45179c74ce 100644
--- a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmax.ll
@@ -6043,14 +6043,14 @@ define half @flat_agent_atomic_fmax_ret_f16__amdgpu_no_fine_grained_memory(ptr %
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6168,14 +6168,14 @@ define half @flat_agent_atomic_fmax_ret_f16__amdgpu_no_fine_grained_memory(ptr %
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
@@ -6438,14 +6438,14 @@ define half @flat_agent_atomic_fmax_ret_f16__offset12b_pos__amdgpu_no_fine_grain
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6570,14 +6570,14 @@ define half @flat_agent_atomic_fmax_ret_f16__offset12b_pos__amdgpu_no_fine_grain
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
@@ -6847,14 +6847,14 @@ define half @flat_agent_atomic_fmax_ret_f16__offset12b_neg__amdgpu_no_fine_grain
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6980,14 +6980,14 @@ define half @flat_agent_atomic_fmax_ret_f16__offset12b_neg__amdgpu_no_fine_grain
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
@@ -7254,13 +7254,12 @@ define void @flat_agent_atomic_fmax_noret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v3.l, v3.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, v2.h, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7376,13 +7375,12 @@ define void @flat_agent_atomic_fmax_noret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v3.l, v3.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, v2.h, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
@@ -7638,13 +7636,12 @@ define void @flat_agent_atomic_fmax_noret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7767,13 +7764,12 @@ define void @flat_agent_atomic_fmax_noret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
@@ -8036,13 +8032,12 @@ define void @flat_agent_atomic_fmax_noret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8166,13 +8161,12 @@ define void @flat_agent_atomic_fmax_noret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
@@ -8424,11 +8418,11 @@ define half @flat_agent_atomic_fmax_ret_f16__offset12b_pos__align4__amdgpu_no_fi
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v4.l, v4.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8519,11 +8513,11 @@ define half @flat_agent_atomic_fmax_ret_f16__offset12b_pos__align4__amdgpu_no_fi
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v4.l, v4.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
@@ -8728,10 +8722,9 @@ define void @flat_agent_atomic_fmax_noret_f16__offset12b__align4_pos__amdgpu_no_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v4.l, v4.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8820,10 +8813,9 @@ define void @flat_agent_atomic_fmax_noret_f16__offset12b__align4_pos__amdgpu_no_
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v4.l, v4.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
@@ -9035,14 +9027,14 @@ define half @flat_system_atomic_fmax_ret_f16__offset12b_pos__amdgpu_no_fine_grai
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -9169,14 +9161,14 @@ define half @flat_system_atomic_fmax_ret_f16__offset12b_pos__amdgpu_no_fine_grai
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
@@ -9448,13 +9440,12 @@ define void @flat_system_atomic_fmax_noret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -9579,13 +9570,12 @@ define void @flat_system_atomic_fmax_noret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll
index f0083bd23660a..5d26293e7009b 100644
--- a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fmin.ll
@@ -6043,14 +6043,14 @@ define half @flat_agent_atomic_fmin_ret_f16__amdgpu_no_fine_grained_memory(ptr %
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6168,14 +6168,14 @@ define half @flat_agent_atomic_fmin_ret_f16__amdgpu_no_fine_grained_memory(ptr %
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
@@ -6438,14 +6438,14 @@ define half @flat_agent_atomic_fmin_ret_f16__offset12b_pos__amdgpu_no_fine_grain
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6570,14 +6570,14 @@ define half @flat_agent_atomic_fmin_ret_f16__offset12b_pos__amdgpu_no_fine_grain
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
@@ -6847,14 +6847,14 @@ define half @flat_agent_atomic_fmin_ret_f16__offset12b_neg__amdgpu_no_fine_grain
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6980,14 +6980,14 @@ define half @flat_agent_atomic_fmin_ret_f16__offset12b_neg__amdgpu_no_fine_grain
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
@@ -7254,13 +7254,12 @@ define void @flat_agent_atomic_fmin_noret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v3.l, v3.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v3.l, v2.h, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7376,13 +7375,12 @@ define void @flat_agent_atomic_fmin_noret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v3.l, v3.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v3.l, v2.h, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
@@ -7638,13 +7636,12 @@ define void @flat_agent_atomic_fmin_noret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7767,13 +7764,12 @@ define void @flat_agent_atomic_fmin_noret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
@@ -8036,13 +8032,12 @@ define void @flat_agent_atomic_fmin_noret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8166,13 +8161,12 @@ define void @flat_agent_atomic_fmin_noret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
@@ -8424,11 +8418,11 @@ define half @flat_agent_atomic_fmin_ret_f16__offset12b_pos__align4__amdgpu_no_fi
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v4.l, v4.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v3.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8519,11 +8513,11 @@ define half @flat_agent_atomic_fmin_ret_f16__offset12b_pos__align4__amdgpu_no_fi
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v4.l, v4.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v3.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
@@ -8728,10 +8722,9 @@ define void @flat_agent_atomic_fmin_noret_f16__offset12b__align4_pos__amdgpu_no_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v4.l, v4.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v3.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8820,10 +8813,9 @@ define void @flat_agent_atomic_fmin_noret_f16__offset12b__align4_pos__amdgpu_no_
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v4.l, v4.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v3.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
@@ -9035,14 +9027,14 @@ define half @flat_system_atomic_fmin_ret_f16__offset12b_pos__amdgpu_no_fine_grai
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -9169,14 +9161,14 @@ define half @flat_system_atomic_fmin_ret_f16__offset12b_pos__amdgpu_no_fine_grai
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
@@ -9448,13 +9440,12 @@ define void @flat_system_atomic_fmin_noret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -9579,13 +9570,12 @@ define void @flat_system_atomic_fmin_noret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[3:4], v[5:6] glc
diff --git a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll
index 3ee0bb2122abe..d12a7f9731586 100644
--- a/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-atomicrmw-fsub.ll
@@ -5855,13 +5855,12 @@ define half @flat_agent_atomic_fsub_ret_f16(ptr %ptr, half %val) #0 {
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -5974,13 +5973,12 @@ define half @flat_agent_atomic_fsub_ret_f16(ptr %ptr, half %val) #0 {
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
@@ -6230,13 +6228,12 @@ define half @flat_agent_atomic_fsub_ret_f16__offset12b_pos(ptr %ptr, half %val)
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6353,13 +6350,12 @@ define half @flat_agent_atomic_fsub_ret_f16__offset12b_pos(ptr %ptr, half %val)
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
@@ -6616,13 +6612,12 @@ define half @flat_agent_atomic_fsub_ret_f16__offset12b_neg(ptr %ptr, half %val)
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6740,13 +6735,12 @@ define half @flat_agent_atomic_fsub_ret_f16__offset12b_neg(ptr %ptr, half %val)
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
@@ -7002,11 +6996,11 @@ define void @flat_agent_atomic_fsub_noret_f16(ptr %ptr, half %val) #0 {
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7116,11 +7110,11 @@ define void @flat_agent_atomic_fsub_noret_f16(ptr %ptr, half %val) #0 {
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
@@ -7362,11 +7356,11 @@ define void @flat_agent_atomic_fsub_noret_f16__offset12b_pos(ptr %ptr, half %val
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7480,11 +7474,11 @@ define void @flat_agent_atomic_fsub_noret_f16__offset12b_pos(ptr %ptr, half %val
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
@@ -7733,11 +7727,11 @@ define void @flat_agent_atomic_fsub_noret_f16__offset12b_neg(ptr %ptr, half %val
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7852,11 +7846,11 @@ define void @flat_agent_atomic_fsub_noret_f16__offset12b_neg(ptr %ptr, half %val
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
@@ -8096,10 +8090,9 @@ define half @flat_agent_atomic_fsub_ret_f16__offset12b_pos__align4(ptr %ptr, hal
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v4.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8185,10 +8178,9 @@ define half @flat_agent_atomic_fsub_ret_f16__offset12b_pos__align4(ptr %ptr, hal
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
@@ -8382,8 +8374,8 @@ define void @flat_agent_atomic_fsub_noret_f16__offset12b__align4_pos(ptr %ptr, h
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v4.l, v2.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8466,8 +8458,8 @@ define void @flat_agent_atomic_fsub_noret_f16__offset12b__align4_pos(ptr %ptr, h
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v4.l, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] offset:2046 glc
@@ -8665,13 +8657,12 @@ define half @flat_system_atomic_fsub_ret_f16__offset12b_pos(ptr %ptr, half %val)
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -8790,13 +8781,12 @@ define half @flat_system_atomic_fsub_ret_f16__offset12b_pos(ptr %ptr, half %val)
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v5, v[0:1], v[5:6] glc
@@ -9055,11 +9045,11 @@ define void @flat_system_atomic_fsub_noret_f16__offset12b_pos(ptr %ptr, half %va
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -9175,11 +9165,11 @@ define void @flat_system_atomic_fsub_noret_f16__offset12b_pos(ptr %ptr, half %va
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: flat_atomic_cmpswap_b32 v3, v[0:1], v[3:4] glc
diff --git a/llvm/test/CodeGen/AMDGPU/fmul-to-ldexp.ll b/llvm/test/CodeGen/AMDGPU/fmul-to-ldexp.ll
index 9c4901eb19f37..899cc89405440 100644
--- a/llvm/test/CodeGen/AMDGPU/fmul-to-ldexp.ll
+++ b/llvm/test/CodeGen/AMDGPU/fmul-to-ldexp.ll
@@ -4238,7 +4238,7 @@ define amdgpu_ps i32 @s_mul_32_f16(half inreg %x, half inreg %y) {
; GFX11-GISEL-TRUE16-LABEL: s_mul_32_f16:
; GFX11-GISEL-TRUE16: ; %bb.0:
; GFX11-GISEL-TRUE16-NEXT: v_mul_f16_e64 v0.l, 0x5000, s0
-; GFX11-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-GISEL-TRUE16-NEXT: v_readfirstlane_b32 s0, v0
; GFX11-GISEL-TRUE16-NEXT: ; return to shader part epilog
;
diff --git a/llvm/test/CodeGen/AMDGPU/fold-int-pow2-with-fmul-or-fdiv.ll b/llvm/test/CodeGen/AMDGPU/fold-int-pow2-with-fmul-or-fdiv.ll
index f09c25767648f..a859cc91b7fde 100644
--- a/llvm/test/CodeGen/AMDGPU/fold-int-pow2-with-fmul-or-fdiv.ll
+++ b/llvm/test/CodeGen/AMDGPU/fold-int-pow2-with-fmul-or-fdiv.ll
@@ -644,11 +644,10 @@ define double @fmul_pow_mul_max_pow2(i16 %cnt) nounwind {
; GFX11-TRUE16-LABEL: fmul_pow_mul_max_pow2:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, v0.l, 2
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-TRUE16-NEXT: v_cvt_f64_u32_e32 v[0:1], v0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_mul_f64 v[0:1], 0x40080000, v[0:1]
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -1194,13 +1193,12 @@ define double @fmul_pow_shl_cnt_safe(i16 %cnt) nounwind {
; GFX11-TRUE16-LABEL: fmul_pow_shl_cnt_safe:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, v0.l, 1
; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0xff5f3992
; GFX11-TRUE16-NEXT: s_mov_b32 s1, 0x7befffff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-TRUE16-NEXT: v_cvt_f64_u32_e32 v[0:1], v0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_mul_f64 v[0:1], v[0:1], s[0:1]
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
diff --git a/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll b/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll
index c52fb6197e3e3..40d2765395543 100644
--- a/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fptrunc.f16.ll
@@ -4372,14 +4372,13 @@ define amdgpu_kernel void @fptrunc_f32_to_f16_zext_i32(
; GFX11-GISEL-TRUE16-LABEL: fptrunc_f32_to_f16_zext_i32:
; GFX11-GISEL-TRUE16: ; %bb.0: ; %entry
; GFX11-GISEL-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
+; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-GISEL-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-TRUE16-NEXT: s_load_b32 s2, s[2:3], 0x0
; GFX11-GISEL-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
; GFX11-GISEL-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-TRUE16-NEXT: v_cvt_f16_f32_e32 v0.l, s2
; GFX11-GISEL-TRUE16-NEXT: s_mov_b32 s2, -1
-; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-GISEL-TRUE16-NEXT: buffer_store_b32 v0, off, s[0:3], 0
; GFX11-GISEL-TRUE16-NEXT: s_endpgm
;
@@ -4607,14 +4606,13 @@ define amdgpu_kernel void @fptrunc_fabs_f32_to_f16_zext_i32(
; GFX11-GISEL-TRUE16-LABEL: fptrunc_fabs_f32_to_f16_zext_i32:
; GFX11-GISEL-TRUE16: ; %bb.0: ; %entry
; GFX11-GISEL-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24
+; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-GISEL-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-TRUE16-NEXT: s_load_b32 s2, s[2:3], 0x0
; GFX11-GISEL-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
; GFX11-GISEL-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-TRUE16-NEXT: v_cvt_f16_f32_e64 v0.l, |s2|
; GFX11-GISEL-TRUE16-NEXT: s_mov_b32 s2, -1
-; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-GISEL-TRUE16-NEXT: buffer_store_b32 v0, off, s[0:3], 0
; GFX11-GISEL-TRUE16-NEXT: s_endpgm
;
diff --git a/llvm/test/CodeGen/AMDGPU/function-args.ll b/llvm/test/CodeGen/AMDGPU/function-args.ll
index 95e28a37f5ee1..3c41cc43a089e 100644
--- a/llvm/test/CodeGen/AMDGPU/function-args.ll
+++ b/llvm/test/CodeGen/AMDGPU/function-args.ll
@@ -1107,21 +1107,19 @@ define void @void_func_v4i8(<4 x i8> %arg0) #0 {
; GFX11-TRUE16-LABEL: void_func_v4i8:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0
; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.h, v0.h
; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l
; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2
; GFX11-TRUE16-NEXT: buffer_store_b32 v0, off, s[0:3], 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -1190,22 +1188,20 @@ define void @void_func_v5i8(<5 x i8> %arg0) #0 {
; GFX11-TRUE16-LABEL: void_func_v5i8:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 4
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l
; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v1.l
; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
; GFX11-TRUE16-NEXT: buffer_store_b8 v4, off, s[0:3], 0
; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v2
; GFX11-TRUE16-NEXT: buffer_store_b32 v0, off, s[0:3], 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -1285,29 +1281,27 @@ define void @void_func_v8i8(<8 x i8> %arg0) #0 {
; GFX11-TRUE16-LABEL: void_func_v8i8:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v5.h, v4.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v6.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h
; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v0.h, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v6
+; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v6.l
; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v4
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v0, v4
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v0, v6
; GFX11-TRUE16-NEXT: buffer_store_b64 v[1:2], off, s[0:3], 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -1422,47 +1416,44 @@ define void @void_func_v16i8(<16 x i8> %arg0) #0 {
; GFX11-TRUE16-LABEL: void_func_v16i8:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l
; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, 0
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.h, v12.h
; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v8.h, v13.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v11.l
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, 0
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v13.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v14.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v10.l, v9.h
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.h, v9.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v14.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l
; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v8.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v5.h, v4.h
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v5.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v14.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v7.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v9, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v10.l, v8.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v11
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v8, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.h, v0.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0
; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v4, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v4, v14
+; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v14.l
; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v0, v12
-; GFX11-TRUE16-NEXT: buffer_store_b128 v[6:9], off, s[0:3], 0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v0, v14
+; GFX11-TRUE16-NEXT: buffer_store_b128 v[5:8], off, s[0:3], 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: void_func_v16i8:
@@ -1658,83 +1649,77 @@ define void @void_func_v32i8(<32 x i8> %arg0) #0 {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: scratch_load_d16_u8 v31, off, s32
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, 0
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v15.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v2.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, 0
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v3.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v32.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l
; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v7.h, v6.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l
; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v1.h, v0.h
-; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v13
+; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v3.h, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v32.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v7.l
; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v29.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v13, v32
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v3.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v6.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v12, v32
; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v5.h, v4.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v25.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff, v14
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v12, v32
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v4.l, v5.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v32
; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v6.l, v7.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v11.h, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v13, v32
-; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v8.h, v8.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v27.l
-; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v17.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v0.h, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v29.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v28.l
+; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v32
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v0.h, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v32.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v27.l
+; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v26.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v25.l
+; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v24.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v6.h, v5.h
; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v32
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.h, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v32.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v23.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v20.l
+; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v7.h, v6.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v32.l
+; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v16.l
; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 16
+; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.h, v5.h
; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000
; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v31.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v9.l, v5.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v23.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v19.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v7, v32
-; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v10.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v18.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v9, v32
-; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v4.h, v5.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v10
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v11, v32
-; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v4.l, v8.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v31.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v4.h, v7.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v13, v32
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v5.l, v4.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v19.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v17.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v14, v32
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v4.h, v8.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v5.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v32
+; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v4.h, v4.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v9, v32
; GFX11-TRUE16-NEXT: buffer_store_b128 v[4:7], off, s[0:3], 0
; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0
diff --git a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
index 2fdc1a8854863..919464a936740 100644
--- a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
+++ b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll
@@ -4896,23 +4896,22 @@ define amdgpu_gfx void @test_call_external_void_func_v4i8_ret() #0 {
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 16, v0
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, 24, v0
; GFX11-TRUE16-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l
; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v42, 1
; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v42, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, 0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2
; GFX11-TRUE16-NEXT: s_mov_b32 s32, s33
; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-TRUE16-NEXT: global_store_b32 v[40:41], v0, off
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s33
@@ -5156,30 +5155,29 @@ define amdgpu_gfx void @test_call_external_void_func_v5i8_ret() #0 {
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, v5
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v6
; GFX11-TRUE16-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v0.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, 4
-; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v42, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_and_b32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v42, 0
-; GFX11-TRUE16-NEXT: s_mov_b32 s32, s33
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v3
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, 0
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: global_store_b8 v[0:1], v4, off
; GFX11-TRUE16-NEXT: global_store_b32 v[40:41], v2, off
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s33
; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s33 offset:4
+; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v42, 1
+; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v42, 0
+; GFX11-TRUE16-NEXT: s_mov_b32 s32, s33
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2
; GFX11-TRUE16-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s33 offset:8 ; 4-byte Folded Reload
; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s1
@@ -5441,36 +5439,34 @@ define amdgpu_gfx void @test_call_external_void_func_v8i8_ret() #0 {
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v7, 24, v1
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v1, v8
; GFX11-TRUE16-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v5.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.h, v0.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v1.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v7.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v7.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.h, v0.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v1.h, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v4.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v4
+; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v4.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v0, v4
; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v42, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v0, v5
; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v42, 0
-; GFX11-TRUE16-NEXT: s_mov_b32 s32, s33
-; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2
; GFX11-TRUE16-NEXT: global_store_b64 v[40:41], v[1:2], off
; GFX11-TRUE16-NEXT: s_clause 0x1
; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s33
; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s33 offset:4
+; GFX11-TRUE16-NEXT: s_mov_b32 s32, s33
+; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2
; GFX11-TRUE16-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s33 offset:8 ; 4-byte Folded Reload
; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s1
@@ -5910,85 +5906,77 @@ define amdgpu_gfx void @test_call_external_void_func_v32i8_ret() #0 {
; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v32 :: v_dual_mov_b32 v18, v33
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v19, v34
; GFX11-TRUE16-NEXT: s_swappc_b64 s[30:31], s[0:1]
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v13.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v12.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v15.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, 0
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v15.l
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v14.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v13.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v12.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, 0
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v1.h, v0.h
-; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v14.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v9.l
-; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff, v12
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v0.h, v2.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v3.h, v1.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v5.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.h, v0.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v11.l
+; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v3.h, v2.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v12.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v10.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v9, v13
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff, v8
-; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v9.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v8.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v12.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v13, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.h, v0.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v7.l
+; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v3.h, v2.h
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v8, v13
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v1.h, v0.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v5.l
+; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v12.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.h, v0.h
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v29.l
+; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v2.h
+; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v12.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v4, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.h, v0.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v31.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v30.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v29.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v28.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v6, v13
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v1.l, v0.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v2, v12
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v17.l
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v0.h, v0.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v27.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v1.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v26.l
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v25.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v24.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v31.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v30.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v13
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v1.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v26.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v6.l, v0.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v27.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v0.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v21.l
+; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v16.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v2, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v0.h, v0.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v23.l
+; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v1.h, v1.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v22.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v21.l
; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v20.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v6, v13
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff, v7
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v22.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v0.l
-; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.h, v1.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v7, v13
-; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v18.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v6.l, v0.h
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v0.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v6, v13
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v1.l, v0.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v10
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v0, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v12
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v0.h, v0.l
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v19.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v12.l
+; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v18.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v12
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v0.h, v0.l
+; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v4.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v12.l
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v12
; GFX11-TRUE16-NEXT: s_clause 0x1
-; GFX11-TRUE16-NEXT: global_store_b128 v[42:43], v[6:9], off
-; GFX11-TRUE16-NEXT: global_store_b128 v[40:41], v[2:5], off
+; GFX11-TRUE16-NEXT: global_store_b128 v[42:43], v[0:3], off
+; GFX11-TRUE16-NEXT: global_store_b128 v[40:41], v[5:8], off
; GFX11-TRUE16-NEXT: s_clause 0x3
; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s33
; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s33 offset:4
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll
index 1f74fbdc46e98..9c1f9d21b9da3 100644
--- a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll
+++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fadd.ll
@@ -8275,13 +8275,12 @@ define half @global_agent_atomic_fadd_ret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8394,13 +8393,12 @@ define half @global_agent_atomic_fadd_ret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
@@ -8700,13 +8698,12 @@ define half @global_agent_atomic_fadd_ret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8823,13 +8820,12 @@ define half @global_agent_atomic_fadd_ret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
@@ -9138,13 +9134,12 @@ define half @global_agent_atomic_fadd_ret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -9262,13 +9257,12 @@ define half @global_agent_atomic_fadd_ret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
@@ -9576,11 +9570,11 @@ define void @global_agent_atomic_fadd_noret_f16__amdgpu_no_fine_grained_memory(p
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -9690,11 +9684,11 @@ define void @global_agent_atomic_fadd_noret_f16__amdgpu_no_fine_grained_memory(p
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
@@ -9985,11 +9979,11 @@ define void @global_agent_atomic_fadd_noret_f16__offset12b_pos__amdgpu_no_fine_g
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -10103,11 +10097,11 @@ define void @global_agent_atomic_fadd_noret_f16__offset12b_pos__amdgpu_no_fine_g
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
@@ -10406,11 +10400,11 @@ define void @global_agent_atomic_fadd_noret_f16__offset12b_neg__amdgpu_no_fine_g
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -10525,11 +10519,11 @@ define void @global_agent_atomic_fadd_noret_f16__offset12b_neg__amdgpu_no_fine_g
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
@@ -10819,10 +10813,9 @@ define half @global_agent_atomic_fadd_ret_f16__offset12b_pos__align4__amdgpu_no_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v4.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -10908,10 +10901,9 @@ define half @global_agent_atomic_fadd_ret_f16__offset12b_pos__align4__amdgpu_no_
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
@@ -11144,8 +11136,8 @@ define void @global_agent_atomic_fadd_noret_f16__offset12b__align4_pos__amdgpu_n
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v4.l, v2.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -11228,8 +11220,8 @@ define void @global_agent_atomic_fadd_noret_f16__offset12b__align4_pos__amdgpu_n
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v4.l, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
@@ -11464,13 +11456,12 @@ define half @global_system_atomic_fadd_ret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -11589,13 +11580,12 @@ define half @global_system_atomic_fadd_ret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
@@ -11906,11 +11896,11 @@ define void @global_system_atomic_fadd_noret_f16__offset12b_pos__amdgpu_no_fine_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -12026,11 +12016,11 @@ define void @global_system_atomic_fadd_noret_f16__offset12b_pos__amdgpu_no_fine_
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll
index faa74fef2be2f..f7cc0709109f9 100644
--- a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll
+++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmax.ll
@@ -4467,14 +4467,14 @@ define half @global_agent_atomic_fmax_ret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -4592,14 +4592,14 @@ define half @global_agent_atomic_fmax_ret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
@@ -4912,14 +4912,14 @@ define half @global_agent_atomic_fmax_ret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -5044,14 +5044,14 @@ define half @global_agent_atomic_fmax_ret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
@@ -5373,14 +5373,14 @@ define half @global_agent_atomic_fmax_ret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -5506,14 +5506,14 @@ define half @global_agent_atomic_fmax_ret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
@@ -5832,13 +5832,12 @@ define void @global_agent_atomic_fmax_noret_f16__amdgpu_no_fine_grained_memory(p
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v3.l, v3.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, v2.h, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -5954,13 +5953,12 @@ define void @global_agent_atomic_fmax_noret_f16__amdgpu_no_fine_grained_memory(p
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v3.l, v3.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, v2.h, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
@@ -6265,13 +6263,12 @@ define void @global_agent_atomic_fmax_noret_f16__offset12b_pos__amdgpu_no_fine_g
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6394,13 +6391,12 @@ define void @global_agent_atomic_fmax_noret_f16__offset12b_pos__amdgpu_no_fine_g
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
@@ -6713,13 +6709,12 @@ define void @global_agent_atomic_fmax_noret_f16__offset12b_neg__amdgpu_no_fine_g
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6843,13 +6838,12 @@ define void @global_agent_atomic_fmax_noret_f16__offset12b_neg__amdgpu_no_fine_g
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
@@ -7151,11 +7145,11 @@ define half @global_agent_atomic_fmax_ret_f16__offset12b_pos__align4__amdgpu_no_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v4.l, v4.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7246,11 +7240,11 @@ define half @global_agent_atomic_fmax_ret_f16__offset12b_pos__align4__amdgpu_no_
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v4.l, v4.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
@@ -7494,10 +7488,9 @@ define void @global_agent_atomic_fmax_noret_f16__offset12b__align4_pos__amdgpu_n
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v4.l, v4.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7586,10 +7579,9 @@ define void @global_agent_atomic_fmax_noret_f16__offset12b__align4_pos__amdgpu_n
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v4.l, v4.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
@@ -7838,14 +7830,14 @@ define half @global_system_atomic_fmax_ret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -7972,14 +7964,14 @@ define half @global_system_atomic_fmax_ret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
@@ -8303,13 +8295,12 @@ define void @global_system_atomic_fmax_noret_f16__offset12b_pos__amdgpu_no_fine_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -8434,13 +8425,12 @@ define void @global_system_atomic_fmax_noret_f16__offset12b_pos__amdgpu_no_fine_
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll
index a46b0129b79e6..b81af1fc9233d 100644
--- a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll
+++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fmin.ll
@@ -4467,14 +4467,14 @@ define half @global_agent_atomic_fmin_ret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -4592,14 +4592,14 @@ define half @global_agent_atomic_fmin_ret_f16__amdgpu_no_fine_grained_memory(ptr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
@@ -4912,14 +4912,14 @@ define half @global_agent_atomic_fmin_ret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -5044,14 +5044,14 @@ define half @global_agent_atomic_fmin_ret_f16__offset12b_pos__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
@@ -5373,14 +5373,14 @@ define half @global_agent_atomic_fmin_ret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -5506,14 +5506,14 @@ define half @global_agent_atomic_fmin_ret_f16__offset12b_neg__amdgpu_no_fine_gra
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
@@ -5832,13 +5832,12 @@ define void @global_agent_atomic_fmin_noret_f16__amdgpu_no_fine_grained_memory(p
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v3.l, v3.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v3.l, v2.h, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -5954,13 +5953,12 @@ define void @global_agent_atomic_fmin_noret_f16__amdgpu_no_fine_grained_memory(p
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v3.l, v3.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v3.l, v2.h, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
@@ -6265,13 +6263,12 @@ define void @global_agent_atomic_fmin_noret_f16__offset12b_pos__amdgpu_no_fine_g
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6394,13 +6391,12 @@ define void @global_agent_atomic_fmin_noret_f16__offset12b_pos__amdgpu_no_fine_g
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
@@ -6713,13 +6709,12 @@ define void @global_agent_atomic_fmin_noret_f16__offset12b_neg__amdgpu_no_fine_g
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6843,13 +6838,12 @@ define void @global_agent_atomic_fmin_noret_f16__offset12b_neg__amdgpu_no_fine_g
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
@@ -7151,11 +7145,11 @@ define half @global_agent_atomic_fmin_ret_f16__offset12b_pos__align4__amdgpu_no_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v4.l, v4.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v3.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7246,11 +7240,11 @@ define half @global_agent_atomic_fmin_ret_f16__offset12b_pos__align4__amdgpu_no_
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v4.l, v4.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v3.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
@@ -7494,10 +7488,9 @@ define void @global_agent_atomic_fmin_noret_f16__offset12b__align4_pos__amdgpu_n
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.h, v4.l, v4.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v3.l, v2.h, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7586,10 +7579,9 @@ define void @global_agent_atomic_fmin_noret_f16__offset12b__align4_pos__amdgpu_n
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.h, v4.l, v4.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v3.l, v2.h, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
@@ -7838,14 +7830,14 @@ define half @global_system_atomic_fmin_ret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -7972,14 +7964,14 @@ define half @global_system_atomic_fmin_ret_f16__offset12b_pos__amdgpu_no_fine_gr
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
@@ -8303,13 +8295,12 @@ define void @global_system_atomic_fmin_noret_f16__offset12b_pos__amdgpu_no_fine_
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v0.h, v5.l, v5.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v5.l, v0.h, v0.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -8434,13 +8425,12 @@ define void @global_system_atomic_fmin_noret_f16__offset12b_pos__amdgpu_no_fine_
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v1, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v0.h, v5.l, v5.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v5.l, v0.h, v0.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v1, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v2, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[3:4], v[5:6], off glc
diff --git a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
index 053efdcb76261..b8762d13e1327 100644
--- a/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
+++ b/llvm/test/CodeGen/AMDGPU/global-atomicrmw-fsub.ll
@@ -5221,13 +5221,12 @@ define half @global_agent_atomic_fsub_ret_f16(ptr addrspace(1) %ptr, half %val)
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -5340,13 +5339,12 @@ define half @global_agent_atomic_fsub_ret_f16(ptr addrspace(1) %ptr, half %val)
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
@@ -5646,13 +5644,12 @@ define half @global_agent_atomic_fsub_ret_f16__offset12b_pos(ptr addrspace(1) %p
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -5769,13 +5766,12 @@ define half @global_agent_atomic_fsub_ret_f16__offset12b_pos(ptr addrspace(1) %p
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
@@ -6084,13 +6080,12 @@ define half @global_agent_atomic_fsub_ret_f16__offset12b_neg(ptr addrspace(1) %p
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6208,13 +6203,12 @@ define half @global_agent_atomic_fsub_ret_f16__offset12b_neg(ptr addrspace(1) %p
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
@@ -6522,11 +6516,11 @@ define void @global_agent_atomic_fsub_noret_f16(ptr addrspace(1) %ptr, half %val
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -6636,11 +6630,11 @@ define void @global_agent_atomic_fsub_noret_f16(ptr addrspace(1) %ptr, half %val
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
@@ -6931,11 +6925,11 @@ define void @global_agent_atomic_fsub_noret_f16__offset12b_pos(ptr addrspace(1)
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7049,11 +7043,11 @@ define void @global_agent_atomic_fsub_noret_f16__offset12b_pos(ptr addrspace(1)
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
@@ -7352,11 +7346,11 @@ define void @global_agent_atomic_fsub_noret_f16__offset12b_neg(ptr addrspace(1)
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7471,11 +7465,11 @@ define void @global_agent_atomic_fsub_noret_f16__offset12b_neg(ptr addrspace(1)
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
@@ -7765,10 +7759,9 @@ define half @global_agent_atomic_fsub_ret_f16__offset12b_pos__align4(ptr addrspa
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v4.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -7854,10 +7847,9 @@ define half @global_agent_atomic_fsub_ret_f16__offset12b_pos__align4(ptr addrspa
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v4.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
@@ -8090,8 +8082,8 @@ define void @global_agent_atomic_fsub_noret_f16__offset12b__align4_pos(ptr addrs
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v4.l, v2.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
@@ -8174,8 +8166,8 @@ define void @global_agent_atomic_fsub_noret_f16__offset12b__align4_pos(ptr addrs
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v4.l, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, 0xffff0000, v4, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off offset:2046 glc
@@ -8410,13 +8402,12 @@ define half @global_system_atomic_fsub_ret_f16__offset12b_pos(ptr addrspace(1) %
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -8535,13 +8526,12 @@ define half @global_system_atomic_fsub_ret_f16__offset12b_pos(ptr addrspace(1) %
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v5, v3, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v5.l, v5.l, v2.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff, v5
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, v3, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v5, v6, v4, v5
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v5, v[0:1], v[5:6], off glc
@@ -8852,11 +8842,11 @@ define void @global_system_atomic_fsub_noret_f16__offset12b_pos(ptr addrspace(1)
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_loadcnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX12-TRUE16-NEXT: global_wb scope:SCOPE_SYS
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
@@ -8972,11 +8962,11 @@ define void @global_system_atomic_fsub_noret_f16__offset12b_pos(ptr addrspace(1)
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v5, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_sub_f16_e32 v3.l, v3.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v5, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v6, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: global_atomic_cmpswap_b32 v3, v[0:1], v[3:4], off glc
diff --git a/llvm/test/CodeGen/AMDGPU/idot4u.ll b/llvm/test/CodeGen/AMDGPU/idot4u.ll
index 7ebd69204d87f..305461ed6b208 100644
--- a/llvm/test/CodeGen/AMDGPU/idot4u.ll
+++ b/llvm/test/CodeGen/AMDGPU/idot4u.ll
@@ -1693,12 +1693,11 @@ define amdgpu_kernel void @notdot4_mixedtypes(ptr addrspace(1) %src1,
; GFX11-DL-TRUE16-NEXT: v_mov_b16_e32 v3.l, v7.l
; GFX11-DL-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-DL-TRUE16-NEXT: v_mad_u16 v0.l, v0.h, v1.l, v0.l
+; GFX11-DL-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-DL-TRUE16-NEXT: v_perm_b32 v1, v5, v5, 0xc0c0302
-; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-DL-TRUE16-NEXT: v_mad_u16 v0.l, v2.l, v3.l, v0.l
; GFX11-DL-TRUE16-NEXT: v_perm_b32 v2, v4, v4, 0xc0c0302
-; GFX11-DL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-DL-TRUE16-NEXT: v_dot4_u32_u8 v0, v2, v1, v0
; GFX11-DL-TRUE16-NEXT: global_store_b16 v6, v0, s[4:5]
; GFX11-DL-TRUE16-NEXT: s_endpgm
@@ -2724,32 +2723,32 @@ define amdgpu_kernel void @udot4_acc8_vecMul(ptr addrspace(1) %src1,
; GFX11-DL-TRUE16-NEXT: global_load_b32 v4, v0, s[2:3]
; GFX11-DL-TRUE16-NEXT: global_load_d16_u8 v0, v5, s[4:5]
; GFX11-DL-TRUE16-NEXT: s_waitcnt vmcnt(2)
-; GFX11-DL-TRUE16-NEXT: v_lshrrev_b16 v0.h, 8, v3.l
-; GFX11-DL-TRUE16-NEXT: s_waitcnt vmcnt(1)
-; GFX11-DL-TRUE16-NEXT: v_lshrrev_b16 v1.l, 8, v4.l
; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 24, v3
+; GFX11-DL-TRUE16-NEXT: s_waitcnt vmcnt(1)
; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v6, 24, v4
+; GFX11-DL-TRUE16-NEXT: v_lshrrev_b16 v0.h, 8, v3.l
+; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v1.l, v3.h, v4.h
+; GFX11-DL-TRUE16-NEXT: v_lshrrev_b16 v1.h, 8, v4.l
; GFX11-DL-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-DL-TRUE16-NEXT: v_mad_u16 v0.l, v3.l, v4.l, v0.l
-; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v0.h, v0.h, v1.l
-; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v1.l, v3.h, v4.h
-; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v1.h, v2.l, v6.l
+; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v2.l, v2.l, v6.l
; GFX11-DL-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0
+; GFX11-DL-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
+; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v0.h, v0.h, v1.h
; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-DL-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v0.h
-; GFX11-DL-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v1.l
-; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-DL-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.h
-; GFX11-DL-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-DL-TRUE16-NEXT: v_or_b16 v6.h, v0.h, v1.l
-; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
-; GFX11-DL-TRUE16-NEXT: v_or_b32_e32 v2, v2, v6
-; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v2, 8, v2
-; GFX11-DL-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v2.l
+; GFX11-DL-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v2.l
+; GFX11-DL-TRUE16-NEXT: v_mov_b16_e32 v7.h, v6.l
+; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-DL-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v0.h
+; GFX11-DL-TRUE16-NEXT: v_or_b16 v6.h, v1.l, v2.l
; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-DL-TRUE16-NEXT: v_or_b32_e32 v1, v7, v6
+; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
+; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-DL-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
+; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v2
; GFX11-DL-TRUE16-NEXT: v_mad_u16 v0.l, v3.h, v4.h, v0.l
+; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-DL-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
; GFX11-DL-TRUE16-NEXT: global_store_b8 v5, v0, s[4:5]
; GFX11-DL-TRUE16-NEXT: s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/integer-mad-patterns.ll b/llvm/test/CodeGen/AMDGPU/integer-mad-patterns.ll
index 742d87f099ce4..31b6b533866d4 100644
--- a/llvm/test/CodeGen/AMDGPU/integer-mad-patterns.ll
+++ b/llvm/test/CodeGen/AMDGPU/integer-mad-patterns.ll
@@ -1715,9 +1715,9 @@ define zeroext i16 @clpeak_umad_pat_i16(i16 zeroext %x, i16 zeroext %y) {
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-TRUE16-NEXT: v_mad_u16 v0.h, v0.l, v1.l, v0.l
; GFX11-SDAG-TRUE16-NEXT: v_mad_u16 v0.l, v0.h, v0.l, v0.h
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-TRUE16-NEXT: v_mad_u16 v0.l, v0.l, v0.h, v0.l
-; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-FAKE16-LABEL: clpeak_umad_pat_i16:
@@ -1745,8 +1745,7 @@ define zeroext i16 @clpeak_umad_pat_i16(i16 zeroext %x, i16 zeroext %y) {
; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-TRUE16-NEXT: v_mul_lo_u16 v0.l, v0.l, v1.l
; GFX11-GISEL-TRUE16-NEXT: v_mul_lo_u16 v0.l, v0.l, v0.h
-; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-FAKE16-LABEL: clpeak_umad_pat_i16:
@@ -1777,9 +1776,9 @@ define zeroext i16 @clpeak_umad_pat_i16(i16 zeroext %x, i16 zeroext %y) {
; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1200-SDAG-TRUE16-NEXT: v_mad_u16 v0.h, v0.l, v1.l, v0.l
; GFX1200-SDAG-TRUE16-NEXT: v_mad_u16 v0.l, v0.h, v0.l, v0.h
-; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1200-SDAG-TRUE16-NEXT: v_mad_u16 v0.l, v0.l, v0.h, v0.l
-; GFX1200-SDAG-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1200-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX1200-SDAG-FAKE16-LABEL: clpeak_umad_pat_i16:
@@ -1815,8 +1814,7 @@ define zeroext i16 @clpeak_umad_pat_i16(i16 zeroext %x, i16 zeroext %y) {
; GFX1200-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1200-GISEL-TRUE16-NEXT: v_mul_lo_u16 v0.l, v0.l, v1.l
; GFX1200-GISEL-TRUE16-NEXT: v_mul_lo_u16 v0.l, v0.l, v0.h
-; GFX1200-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1200-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1200-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1200-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX1200-GISEL-FAKE16-LABEL: clpeak_umad_pat_i16:
@@ -9363,9 +9361,9 @@ define zeroext i16 @clpeak_umad_pat_i16_x2(i16 zeroext %x, i16 zeroext %y) {
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-TRUE16-NEXT: v_mad_u16 v0.h, v0.l, v0.h, v0.l
; GFX11-SDAG-TRUE16-NEXT: v_mad_u16 v0.l, v0.h, v0.l, v0.h
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-TRUE16-NEXT: v_mad_u16 v0.l, v0.l, v0.h, v0.l
-; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-FAKE16-LABEL: clpeak_umad_pat_i16_x2:
@@ -9409,8 +9407,7 @@ define zeroext i16 @clpeak_umad_pat_i16_x2(i16 zeroext %x, i16 zeroext %y) {
; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-TRUE16-NEXT: v_mul_lo_u16 v0.l, v0.l, v1.l
; GFX11-GISEL-TRUE16-NEXT: v_mul_lo_u16 v0.l, v0.l, v0.h
-; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-FAKE16-LABEL: clpeak_umad_pat_i16_x2:
@@ -9457,9 +9454,9 @@ define zeroext i16 @clpeak_umad_pat_i16_x2(i16 zeroext %x, i16 zeroext %y) {
; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1200-SDAG-TRUE16-NEXT: v_mad_u16 v0.h, v0.l, v0.h, v0.l
; GFX1200-SDAG-TRUE16-NEXT: v_mad_u16 v0.l, v0.h, v0.l, v0.h
-; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1200-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX1200-SDAG-TRUE16-NEXT: v_mad_u16 v0.l, v0.l, v0.h, v0.l
-; GFX1200-SDAG-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1200-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1200-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX1200-SDAG-FAKE16-LABEL: clpeak_umad_pat_i16_x2:
@@ -9511,8 +9508,7 @@ define zeroext i16 @clpeak_umad_pat_i16_x2(i16 zeroext %x, i16 zeroext %y) {
; GFX1200-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX1200-GISEL-TRUE16-NEXT: v_mul_lo_u16 v0.l, v0.l, v1.l
; GFX1200-GISEL-TRUE16-NEXT: v_mul_lo_u16 v0.l, v0.l, v0.h
-; GFX1200-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1200-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX1200-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX1200-GISEL-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX1200-GISEL-FAKE16-LABEL: clpeak_umad_pat_i16_x2:
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fadd.ll b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fadd.ll
index a42c71c4849bd..c1a32aafbc71e 100644
--- a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fadd.ll
+++ b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fadd.ll
@@ -1259,13 +1259,12 @@ define half @local_atomic_fadd_ret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, 4.0, v3.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
@@ -1371,13 +1370,12 @@ define half @local_atomic_fadd_ret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, 4.0, v3.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
@@ -1646,13 +1644,12 @@ define half @local_atomic_fadd_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, 4.0, v3.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
@@ -1763,13 +1760,12 @@ define half @local_atomic_fadd_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, 4.0, v3.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
@@ -2044,13 +2040,12 @@ define void @local_atomic_fadd_noret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v4.l, 4.0, v4.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
@@ -2153,13 +2148,12 @@ define void @local_atomic_fadd_noret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v4.l, 4.0, v4.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
@@ -2419,11 +2413,11 @@ define void @local_atomic_fadd_noret_f16__offset(ptr addrspace(3) %ptr) nounwind
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v4.l, 4.0, v4.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
@@ -2531,11 +2525,11 @@ define void @local_atomic_fadd_noret_f16__offset(ptr addrspace(3) %ptr) nounwind
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v4.l, 4.0, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
@@ -2795,10 +2789,9 @@ define half @local_atomic_fadd_ret_f16__offset__align4(ptr addrspace(3) %ptr) no
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v1
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v1.l, 4.0, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
@@ -2882,10 +2875,9 @@ define half @local_atomic_fadd_ret_f16__offset__align4(ptr addrspace(3) %ptr) no
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v1.l, 4.0, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
@@ -3095,8 +3087,8 @@ define void @local_atomic_fadd_noret_f16__offset__align4(ptr addrspace(3) %ptr)
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v2.l, 4.0, v1.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
@@ -3177,8 +3169,8 @@ define void @local_atomic_fadd_noret_f16__offset__align4(ptr addrspace(3) %ptr)
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v2.l, 4.0, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmax.ll b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmax.ll
index 8351d28057564..739e86d1928b1 100644
--- a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmax.ll
+++ b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmax.ll
@@ -803,14 +803,14 @@ define half @local_atomic_fmax_ret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, v3.l, v3.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, 4.0, v3.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
@@ -918,14 +918,14 @@ define half @local_atomic_fmax_ret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, v3.l, v3.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, 4.0, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
@@ -1199,14 +1199,14 @@ define half @local_atomic_fmax_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, v3.l, v3.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, 4.0, v3.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
@@ -1319,14 +1319,14 @@ define half @local_atomic_fmax_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, v3.l, v3.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, 4.0, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
@@ -1606,14 +1606,14 @@ define void @local_atomic_fmax_noret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v4.l, v4.l, v4.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v4.l, 4.0, v4.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
@@ -1718,14 +1718,14 @@ define void @local_atomic_fmax_noret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v4.l, v4.l, v4.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v4.l, 4.0, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
@@ -1990,13 +1990,12 @@ define void @local_atomic_fmax_noret_f16__offset(ptr addrspace(3) %ptr) nounwind
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v4.l, v4.l, v4.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v4.l, 4.0, v4.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
@@ -2107,13 +2106,12 @@ define void @local_atomic_fmax_noret_f16__offset(ptr addrspace(3) %ptr) nounwind
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v4.l, v4.l, v4.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v4.l, 4.0, v4.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
@@ -2379,11 +2377,11 @@ define half @local_atomic_fmax_ret_f16__offset__align4(ptr addrspace(3) %ptr) no
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v1
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v1.l, v2.l, v2.l
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v1.l, 4.0, v1.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
@@ -2469,11 +2467,11 @@ define half @local_atomic_fmax_ret_f16__offset__align4(ptr addrspace(3) %ptr) no
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v1.l, v2.l, v2.l
; GFX11-TRUE16-NEXT: v_max_f16_e32 v1.l, 4.0, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
@@ -2688,10 +2686,9 @@ define void @local_atomic_fmax_noret_f16__offset__align4(ptr addrspace(3) %ptr)
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.l, v1.l, v1.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.l, 4.0, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
@@ -2775,10 +2772,9 @@ define void @local_atomic_fmax_noret_f16__offset__align4(ptr addrspace(3) %ptr)
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.l, v1.l, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.l, 4.0, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll
index 0c4aca88b3781..6da80262951e5 100644
--- a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll
+++ b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fmin.ll
@@ -803,14 +803,14 @@ define half @local_atomic_fmin_ret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, v3.l, v3.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v3.l, 4.0, v3.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
@@ -918,14 +918,14 @@ define half @local_atomic_fmin_ret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, v3.l, v3.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v3.l, 4.0, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
@@ -1199,14 +1199,14 @@ define half @local_atomic_fmin_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v3.l, v3.l, v3.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v3.l, 4.0, v3.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
@@ -1319,14 +1319,14 @@ define half @local_atomic_fmin_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v3.l, v3.l, v3.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v3.l, 4.0, v3.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
@@ -1606,14 +1606,14 @@ define void @local_atomic_fmin_noret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v4.l, v4.l, v4.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v4.l, 4.0, v4.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
@@ -1718,14 +1718,14 @@ define void @local_atomic_fmin_noret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX11-TRUE16-NEXT: v_max_f16_e32 v4.l, v4.l, v4.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v4.l, 4.0, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
@@ -1990,13 +1990,12 @@ define void @local_atomic_fmin_noret_f16__offset(ptr addrspace(3) %ptr) nounwind
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v4.l, v4.l, v4.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v4.l, 4.0, v4.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
@@ -2107,13 +2106,12 @@ define void @local_atomic_fmin_noret_f16__offset(ptr addrspace(3) %ptr) nounwind
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v4.l, v4.l, v4.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v4.l, 4.0, v4.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
@@ -2379,11 +2377,11 @@ define half @local_atomic_fmin_ret_f16__offset__align4(ptr addrspace(3) %ptr) no
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v1
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v1.l, v2.l, v2.l
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v1.l, 4.0, v1.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
@@ -2469,11 +2467,11 @@ define half @local_atomic_fmin_ret_f16__offset__align4(ptr addrspace(3) %ptr) no
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v1.l, v2.l, v2.l
; GFX11-TRUE16-NEXT: v_min_f16_e32 v1.l, 4.0, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
@@ -2688,10 +2686,9 @@ define void @local_atomic_fmin_noret_f16__offset__align4(ptr addrspace(3) %ptr)
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_max_num_f16_e32 v2.l, v1.l, v1.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_min_num_f16_e32 v2.l, 4.0, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
@@ -2775,10 +2772,9 @@ define void @local_atomic_fmin_noret_f16__offset__align4(ptr addrspace(3) %ptr)
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_max_f16_e32 v2.l, v1.l, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_min_f16_e32 v2.l, 4.0, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll
index 37310b614c0db..786989cc9fb57 100644
--- a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll
+++ b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll
@@ -1721,13 +1721,12 @@ define half @local_atomic_fsub_ret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, -4.0, v3.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
@@ -1833,13 +1832,12 @@ define half @local_atomic_fsub_ret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, -4.0, v3.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4
@@ -2108,13 +2106,12 @@ define half @local_atomic_fsub_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, -4.0, v3.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
@@ -2225,13 +2222,12 @@ define half @local_atomic_fsub_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, -4.0, v3.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4
@@ -2506,13 +2502,12 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX12-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v4.l, -4.0, v4.l
; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
@@ -2615,13 +2610,12 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX11-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
; GFX11-TRUE16-NEXT: v_add_f16_e32 v4.l, -4.0, v4.l
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2
@@ -2881,11 +2875,11 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v4.l, -4.0, v4.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
@@ -2993,11 +2987,11 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v4.l, -4.0, v4.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3
@@ -3257,10 +3251,9 @@ define half @local_atomic_fsub_ret_f16__offset__align4(ptr addrspace(3) %ptr) no
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v1
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_add_f16_e32 v1.l, -4.0, v2.l
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
@@ -3344,10 +3337,9 @@ define half @local_atomic_fsub_ret_f16__offset__align4(ptr addrspace(3) %ptr) no
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v1.l, -4.0, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
@@ -3557,8 +3549,8 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr)
; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0
; GFX12-TRUE16-NEXT: v_add_f16_e32 v2.l, -4.0, v1.l
-; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0
; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
@@ -3639,8 +3631,8 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr)
; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1
; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v2.l, -4.0, v1.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
diff --git a/llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll b/llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll
index 811e25587d3d5..eab92668c536b 100644
--- a/llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll
+++ b/llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll
@@ -2382,13 +2382,22 @@ define <4 x half> @v_mad_mix_v4f32_clamp_precvt(<4 x half> %src0, <4 x half> %sr
}
define i32 @mixlo_zext(float %src0, float %src1, float %src2) #0 {
-; GFX1100-LABEL: mixlo_zext:
-; GFX1100: ; %bb.0:
-; GFX1100-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX1100-NEXT: v_fma_mixlo_f16 v0, v0, v1, v2
-; GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX1100-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX1100-NEXT: s_setpc_b64 s[30:31]
+; SDAG-GFX1100-TRUE16-LABEL: mixlo_zext:
+; SDAG-GFX1100-TRUE16: ; %bb.0:
+; SDAG-GFX1100-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SDAG-GFX1100-TRUE16-NEXT: v_fma_mixlo_f16 v1, v0, v1, v2
+; SDAG-GFX1100-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
+; SDAG-GFX1100-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; SDAG-GFX1100-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.l
+; SDAG-GFX1100-TRUE16-NEXT: s_setpc_b64 s[30:31]
+;
+; SDAG-GFX1100-FAKE16-LABEL: mixlo_zext:
+; SDAG-GFX1100-FAKE16: ; %bb.0:
+; SDAG-GFX1100-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SDAG-GFX1100-FAKE16-NEXT: v_fma_mixlo_f16 v0, v0, v1, v2
+; SDAG-GFX1100-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; SDAG-GFX1100-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; SDAG-GFX1100-FAKE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX900-LABEL: mixlo_zext:
; GFX900: ; %bb.0:
@@ -2418,6 +2427,14 @@ define i32 @mixlo_zext(float %src0, float %src1, float %src2) #0 {
; SDAG-CI-NEXT: v_cvt_f16_f32_e32 v0, v2
; SDAG-CI-NEXT: s_setpc_b64 s[30:31]
;
+; GISEL-GFX1100-LABEL: mixlo_zext:
+; GISEL-GFX1100: ; %bb.0:
+; GISEL-GFX1100-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GISEL-GFX1100-NEXT: v_fma_mixlo_f16 v0, v0, v1, v2
+; GISEL-GFX1100-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GISEL-GFX1100-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GISEL-GFX1100-NEXT: s_setpc_b64 s[30:31]
+;
; GISEL-CI-LABEL: mixlo_zext:
; GISEL-CI: ; %bb.0:
; GISEL-CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/mad.u16.ll b/llvm/test/CodeGen/AMDGPU/mad.u16.ll
index ef80323a98ec0..fbf8011fd40c9 100644
--- a/llvm/test/CodeGen/AMDGPU/mad.u16.ll
+++ b/llvm/test/CodeGen/AMDGPU/mad.u16.ll
@@ -179,8 +179,7 @@ define i32 @v_mad_u16_zext(i16 %arg0, i16 %arg1, i16 %arg2) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mad_u16 v0.l, v0.l, v1.l, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_mad_u16_zext:
@@ -222,9 +221,9 @@ define i64 @v_mad_u16_zext64(i16 %arg0, i16 %arg1, i16 %arg2) {
; GFX11-TRUE16-LABEL: v_mad_u16_zext64:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: v_mad_u16 v0.l, v0.l, v1.l, v2.l
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_and_b32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: v_mad_u16_zext64:
diff --git a/llvm/test/CodeGen/AMDGPU/preserve-hi16.ll b/llvm/test/CodeGen/AMDGPU/preserve-hi16.ll
index 3ce09475c0949..79910af5c0434 100644
--- a/llvm/test/CodeGen/AMDGPU/preserve-hi16.ll
+++ b/llvm/test/CodeGen/AMDGPU/preserve-hi16.ll
@@ -374,7 +374,7 @@ define i32 @shl_i16_zext_i32(i16 %x, i16 %y) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: shl_i16_zext_i32:
@@ -412,7 +412,7 @@ define i32 @lshr_i16_zext_i32(i16 %x, i16 %y) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_lshrrev_b16 v0.l, v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: lshr_i16_zext_i32:
@@ -450,7 +450,7 @@ define i32 @ashr_i16_zext_i32(i16 %x, i16 %y) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_ashrrev_i16 v0.l, v1.l, v0.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: ashr_i16_zext_i32:
@@ -488,7 +488,7 @@ define i32 @add_u16_zext_i32(i16 %x, i16 %y) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: add_u16_zext_i32:
@@ -526,7 +526,7 @@ define i32 @sub_u16_zext_i32(i16 %x, i16 %y) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_sub_nc_u16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: sub_u16_zext_i32:
@@ -564,7 +564,7 @@ define i32 @mul_lo_u16_zext_i32(i16 %x, i16 %y) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_mul_lo_u16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: mul_lo_u16_zext_i32:
@@ -602,7 +602,7 @@ define i32 @min_u16_zext_i32(i16 %x, i16 %y) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_min_u16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: min_u16_zext_i32:
@@ -641,7 +641,7 @@ define i32 @min_i16_zext_i32(i16 %x, i16 %y) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_min_i16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: min_i16_zext_i32:
@@ -680,7 +680,7 @@ define i32 @max_u16_zext_i32(i16 %x, i16 %y) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_max_u16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: max_u16_zext_i32:
@@ -719,7 +719,7 @@ define i32 @max_i16_zext_i32(i16 %x, i16 %y) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_max_i16 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: max_i16_zext_i32:
@@ -758,7 +758,7 @@ define i32 @zext_fadd_f16(half %x, half %y) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_add_f16_e32 v0.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: zext_fadd_f16:
@@ -797,8 +797,10 @@ define i32 @zext_fma_f16(half %x, half %y, half %z) {
; GFX11-TRUE16-LABEL: zext_fma_f16:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: v_fmac_f16_e32 v2.l, v0.l, v1.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v0.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.l
+; GFX11-TRUE16-NEXT: v_fmac_f16_e32 v0.l, v0.h, v1.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: zext_fma_f16:
@@ -838,7 +840,7 @@ define i32 @zext_div_fixup_f16(half %x, half %y, half %z) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_div_fixup_f16 v0.l, v0.l, v1.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: zext_div_fixup_f16:
@@ -880,7 +882,7 @@ define i32 @zext_fptrunc_f16(float %x) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_cvt_f16_f32_e32 v0.l, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: zext_fptrunc_f16:
@@ -924,12 +926,20 @@ define i32 @zext_fptrunc_fma_f16(float %x, float %y, float %z) {
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
-; GFX11-LABEL: zext_fptrunc_fma_f16:
-; GFX11: ; %bb.0:
-; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_fma_mixlo_f16 v0, v0, v1, v2
-; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
-; GFX11-NEXT: s_setpc_b64 s[30:31]
+; GFX11-TRUE16-LABEL: zext_fptrunc_fma_f16:
+; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: v_fma_mixlo_f16 v1, v0, v1, v2
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.l
+; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX11-FAKE16-LABEL: zext_fptrunc_fma_f16:
+; GFX11-FAKE16: ; %bb.0:
+; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-FAKE16-NEXT: v_fma_mixlo_f16 v0, v0, v1, v2
+; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
+; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
%fma = call float @llvm.fma.f32(float %x, float %y, float %z)
%fptrunc = fptrunc float %fma to half
%cast = bitcast half %fptrunc to i16
@@ -940,3 +950,5 @@ define i32 @zext_fptrunc_fma_f16(float %x, float %y, float %z) {
declare half @llvm.amdgcn.div.fixup.f16(half, half, half)
declare half @llvm.fma.f16(half, half, half)
declare float @llvm.fma.f32(float, float, float)
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; GFX11: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/shrink-add-sub-constant.ll b/llvm/test/CodeGen/AMDGPU/shrink-add-sub-constant.ll
index 21aa40d69998e..91c88ec5e718c 100644
--- a/llvm/test/CodeGen/AMDGPU/shrink-add-sub-constant.ll
+++ b/llvm/test/CodeGen/AMDGPU/shrink-add-sub-constant.ll
@@ -1528,10 +1528,9 @@ define amdgpu_kernel void @v_test_i16_x_sub_64_zext_to_i32(ptr addrspace(1) %out
; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 2, v1
; GFX11-SDAG-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-TRUE16-NEXT: global_load_d16_b16 v0, v0, s[2:3]
+; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-SDAG-TRUE16-NEXT: v_sub_nc_u16 v0.l, v0.l, 64
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-SDAG-TRUE16-NEXT: global_store_b32 v1, v0, s[0:1]
; GFX11-SDAG-TRUE16-NEXT: s_endpgm
;
@@ -1560,10 +1559,9 @@ define amdgpu_kernel void @v_test_i16_x_sub_64_zext_to_i32(ptr addrspace(1) %out
; GFX11-GISEL-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 2, v1
; GFX11-GISEL-TRUE16-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-TRUE16-NEXT: global_load_d16_b16 v0, v0, s[2:3]
+; GFX11-GISEL-TRUE16-NEXT: v_mov_b16_e32 v0.h, 0
; GFX11-GISEL-TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11-GISEL-TRUE16-NEXT: v_add_nc_u16 v0.l, 0xffc0, v0.l
-; GFX11-GISEL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-GISEL-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-GISEL-TRUE16-NEXT: global_store_b32 v1, v0, s[0:1]
; GFX11-GISEL-TRUE16-NEXT: s_endpgm
;
diff --git a/llvm/test/CodeGen/AMDGPU/vector-reduce-add.ll b/llvm/test/CodeGen/AMDGPU/vector-reduce-add.ll
index 30ed6ae5484c6..334215125f58a 100644
--- a/llvm/test/CodeGen/AMDGPU/vector-reduce-add.ll
+++ b/llvm/test/CodeGen/AMDGPU/vector-reduce-add.ll
@@ -300,17 +300,15 @@ define i8 @test_vector_reduce_add_v4i8(<4 x i8> %v) {
; GFX11-SDAG-TRUE16-LABEL: test_vector_reduce_add_v4i8:
; GFX11-SDAG-TRUE16: ; %bb.0: ; %entry
; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v3.l
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v1.l, v3.l
; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v2.l
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v0.h
; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, v1.l
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX11-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v2
+; GFX11-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -348,17 +346,15 @@ define i8 @test_vector_reduce_add_v4i8(<4 x i8> %v) {
; GFX12-SDAG-TRUE16-NEXT: s_wait_samplecnt 0x0
; GFX12-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0
; GFX12-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v3.l
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v1.l, v3.l
; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v2.l
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
+; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v0.h
; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v2.l, v1.l
+; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX12-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v2
-; GFX12-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v2
+; GFX12-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
; GFX12-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -518,21 +514,19 @@ define i8 @test_vector_reduce_add_v8i8(<8 x i8> %v) {
; GFX11-SDAG-TRUE16-LABEL: test_vector_reduce_add_v8i8:
; GFX11-SDAG-TRUE16: ; %bb.0: ; %entry
; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v2.l, v6.l
; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.h, v3.l, v7.l
; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v5.l
; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v4.l
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v1.l, v1.h
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v2.l, v6.l
-; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v0.h
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
-; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v1.h
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v0.h
+; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
; GFX11-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
@@ -581,21 +575,19 @@ define i8 @test_vector_reduce_add_v8i8(<8 x i8> %v) {
; GFX12-SDAG-TRUE16-NEXT: s_wait_samplecnt 0x0
; GFX12-SDAG-TRUE16-NEXT: s_wait_bvhcnt 0x0
; GFX12-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v2.l, v6.l
; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.h, v3.l, v7.l
; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v5.l
; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v4.l
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v1.l, v1.h
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v2.l, v6.l
-; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v0.h
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
-; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l
; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v1.h
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v0.h
+; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l
+; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX12-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
; GFX12-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
@@ -832,28 +824,25 @@ define i8 @test_vector_reduce_add_v16i8(<16 x i8> %v) {
; GFX11-SDAG-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.h, v5.l, v13.l
; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v9.l
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v3.h, v7.l, v15.l
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v3.l, v3.l, v11.l
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v2.h, v6.l, v14.l
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v5.l, v7.l, v15.l
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v6.l, v14.l
; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v2.l, v2.l, v10.l
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v1.h
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, v12.l
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v3.l, v3.h
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v2.h, v3.l, v11.l
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v3.l, v4.l, v12.l
; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v8.l
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v1.l, v0.h
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v2.l, v2.h
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.h
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v1.h
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v2.l, v0.h
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v2.l, v2.h, v5.l
+; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v3.l
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v0.h
-; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v2.l
+; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v0.h
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l
+; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX11-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
; GFX11-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
; GFX11-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -922,28 +911,25 @@ define i8 @test_vector_reduce_add_v16i8(<16 x i8> %v) {
; GFX12-SDAG-TRUE16-NEXT: s_wait_kmcnt 0x0
; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.h, v5.l, v13.l
; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v9.l
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v3.h, v7.l, v15.l
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v3.l, v3.l, v11.l
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v2.h, v6.l, v14.l
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v5.l, v7.l, v15.l
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v6.l, v14.l
; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v2.l, v2.l, v10.l
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v1.h
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, v12.l
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v3.l, v3.h
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v2.h, v3.l, v11.l
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v3.l, v4.l, v12.l
; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v8.l
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v1.l, v0.h
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v2.l, v2.h
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.h
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v1.h
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.h, v2.l, v0.h
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v2.l, v2.h, v5.l
+; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v3.l
; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v0.h
-; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v1.l, v1.l, v2.l
+; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v0.h
; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v2.l
+; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
-; GFX12-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX12-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l
; GFX12-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
; GFX12-SDAG-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l
; GFX12-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/vector-reduce-umin.ll b/llvm/test/CodeGen/AMDGPU/vector-reduce-umin.ll
index aab0e76410ccb..1d3b42ee43b0f 100644
--- a/llvm/test/CodeGen/AMDGPU/vector-reduce-umin.ll
+++ b/llvm/test/CodeGen/AMDGPU/vector-reduce-umin.ll
@@ -374,13 +374,12 @@ define i8 @test_vector_reduce_umin_v4i8(<4 x i8> %v) {
; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l
; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-SDAG-TRUE16-NEXT: v_min_u16 v1.l, v1.l, v1.h
+; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX11-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-TRUE16-NEXT: v_min3_u16 v0.l, v0.l, v0.h, v1.l
; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -427,13 +426,12 @@ define i8 @test_vector_reduce_umin_v4i8(<4 x i8> %v) {
; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l
; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-SDAG-TRUE16-NEXT: v_min_u16 v1.l, v1.l, v1.h
+; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX12-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-SDAG-TRUE16-NEXT: v_min3_u16 v0.l, v0.l, v0.h, v1.l
; GFX12-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -624,22 +622,20 @@ define i8 @test_vector_reduce_umin_v8i8(<8 x i8> %v) {
; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v7.l
; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l
; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX11-SDAG-TRUE16-NEXT: v_min_u16 v1.l, v1.l, v1.h
-; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
+; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX11-SDAG-TRUE16-NEXT: v_min_u16 v0.h, v1.l, v1.h
+; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l
+; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-SDAG-TRUE16-NEXT: v_min_u16 v0.l, v0.l, v0.h
-; GFX11-SDAG-TRUE16-NEXT: v_min3_u16 v1.l, v1.l, v3.l, v3.h
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
-; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l
-; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
-; GFX11-SDAG-TRUE16-NEXT: v_min3_u16 v0.l, v0.l, v1.h, v1.l
+; GFX11-SDAG-TRUE16-NEXT: v_min3_u16 v0.h, v0.h, v3.l, v3.h
+; GFX11-SDAG-TRUE16-NEXT: v_min_u16 v0.l, v0.l, v1.l
+; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v0.h
+; GFX11-SDAG-TRUE16-NEXT: v_min3_u16 v0.l, v0.l, v2.l, v1.h
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v2
+; GFX11-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v3
; GFX11-SDAG-TRUE16-NEXT: v_min_u16 v0.l, v0.l, v1.l
; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -703,22 +699,20 @@ define i8 @test_vector_reduce_umin_v8i8(<8 x i8> %v) {
; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l
; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v7.l
; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l
-; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l
; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
-; GFX12-SDAG-TRUE16-NEXT: v_min_u16 v1.l, v1.l, v1.h
-; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l
+; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l
+; GFX12-SDAG-TRUE16-NEXT: v_min_u16 v0.h, v1.l, v1.h
+; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l
+; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l
; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX12-SDAG-TRUE16-NEXT: v_min_u16 v0.l, v0.l, v0.h
-; GFX12-SDAG-TRUE16-NEXT: v_min3_u16 v1.l, v1.l, v3.l, v3.h
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l
-; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.l
-; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v3
-; GFX12-SDAG-TRUE16-NEXT: v_min3_u16 v0.l, v0.l, v1.h, v1.l
+; GFX12-SDAG-TRUE16-NEXT: v_min3_u16 v0.h, v0.h, v3.l, v3.h
+; GFX12-SDAG-TRUE16-NEXT: v_min_u16 v0.l, v0.l, v1.l
+; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0
+; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v0.h
+; GFX12-SDAG-TRUE16-NEXT: v_min3_u16 v0.l, v0.l, v2.l, v1.h
; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v2
+; GFX12-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v3
; GFX12-SDAG-TRUE16-NEXT: v_min_u16 v0.l, v0.l, v1.l
; GFX12-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -1047,14 +1041,12 @@ define i8 @test_vector_reduce_umin_v16i8(<16 x i8> %v) {
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-SDAG-TRUE16-NEXT: v_min3_u16 v0.l, v0.l, v1.h, v1.l
; GFX11-SDAG-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v0.h
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h
-; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
-; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX11-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-SDAG-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX11-SDAG-TRUE16-NEXT: v_or_b16 v1.l, v0.l, v0.h
+; GFX11-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.l
+; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
-; GFX11-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-TRUE16-NEXT: v_min_u16 v0.l, v0.l, v1.l
; GFX11-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -1176,14 +1168,12 @@ define i8 @test_vector_reduce_umin_v16i8(<16 x i8> %v) {
; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-SDAG-TRUE16-NEXT: v_min3_u16 v0.l, v0.l, v1.h, v1.l
; GFX12-SDAG-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v0.h
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-SDAG-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v0.h
-; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.l, v0.l
-; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l
+; GFX12-SDAG-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0
; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-SDAG-TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
+; GFX12-SDAG-TRUE16-NEXT: v_or_b16 v1.l, v0.l, v0.h
+; GFX12-SDAG-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v1.l
+; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-SDAG-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1
-; GFX12-SDAG-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-SDAG-TRUE16-NEXT: v_min_u16 v0.l, v0.l, v1.l
; GFX12-SDAG-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
More information about the llvm-commits
mailing list