[llvm] 97a1f49 - DAG: Avoid breaking legal vector_shuffle with multiple uses (#123712)
via llvm-commits
llvm-commits at lists.llvm.org
Wed Jan 29 19:55:25 PST 2025
Author: Matt Arsenault
Date: 2025-01-30T10:55:21+07:00
New Revision: 97a1f494a6778df24cfca817a30b73b1f0a7228a
URL: https://github.com/llvm/llvm-project/commit/97a1f494a6778df24cfca817a30b73b1f0a7228a
DIFF: https://github.com/llvm/llvm-project/commit/97a1f494a6778df24cfca817a30b73b1f0a7228a.diff
LOG: DAG: Avoid breaking legal vector_shuffle with multiple uses (#123712)
Previously this combine would undo AMDGPU's new custom legalization of
wide vector shuffles into 2 element pieces. The comment also
states that this combine is only done before legalization,
but the case with a build_vector source was unconditional.
We probably don't want to do this if the multiple uses are full
scalarization of the vector, but this seems to work well enough.
Scalarizing extracts should have folded out pre-legalize.
Added:
Modified:
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v2f32.ll
llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v3f32.ll
llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v4f32.ll
llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v2i32.ll
llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v3i32.ll
llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v4i32.ll
llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v2p3.ll
llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v3p3.ll
llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v4p3.ll
Removed:
################################################################################
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index a0c703d2df8a2f..f4caaf426de6a0 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -23186,6 +23186,11 @@ SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
}
if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
+ // TODO: Check if shuffle mask is legal?
+ if (LegalOperations && TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VecVT) &&
+ !VecOp.hasOneUse())
+ return SDValue();
+
SDValue InOp = SVInVec.getOperand(OrigElt);
if (InOp.getValueType() != ScalarVT) {
assert(InOp.getValueType().isInteger() && ScalarVT.isInteger());
diff --git a/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v2f32.ll b/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v2f32.ll
index fafdac7ffac62f..b242517b50841d 100644
--- a/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v2f32.ll
+++ b/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v2f32.ll
@@ -176,8 +176,7 @@ define void @v_shuffle_v4f32_v2f32__3_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -192,8 +191,8 @@ define void @v_shuffle_v4f32_v2f32__3_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -566,16 +565,15 @@ define void @v_shuffle_v4f32_v2f32__3_3_3_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:1]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -583,16 +581,15 @@ define void @v_shuffle_v4f32_v2f32__3_3_3_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:1]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x float> asm "; def $0", "=v"()
@@ -675,31 +672,26 @@ define void @v_shuffle_v4f32_v2f32__3_3_3_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4f32_v2f32__3_3_3_2:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v5
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4f32_v2f32__3_3_3_2:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v5
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x float> asm "; def $0", "=v"()
@@ -873,8 +865,7 @@ define void @v_shuffle_v4f32_v2f32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -887,8 +878,7 @@ define void @v_shuffle_v4f32_v2f32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -972,8 +962,7 @@ define void @v_shuffle_v4f32_v2f32__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -989,8 +978,7 @@ define void @v_shuffle_v4f32_v2f32__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -1138,14 +1126,13 @@ define void @v_shuffle_v4f32_v2f32__3_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1155,14 +1142,13 @@ define void @v_shuffle_v4f32_v2f32__3_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x float> asm "; def $0", "=v"()
@@ -1303,16 +1289,15 @@ define void @v_shuffle_v4f32_v2f32__3_3_1_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:1]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1320,16 +1305,15 @@ define void @v_shuffle_v4f32_v2f32__3_3_1_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:1]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x float> asm "; def $0", "=v"()
@@ -1723,8 +1707,7 @@ define void @v_shuffle_v4f32_v2f32__3_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -1740,8 +1723,7 @@ define void @v_shuffle_v4f32_v2f32__3_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -1776,14 +1758,13 @@ define void @v_shuffle_v4f32_v2f32__3_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1793,14 +1774,13 @@ define void @v_shuffle_v4f32_v2f32__3_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x float> asm "; def $0", "=v"()
@@ -2152,8 +2132,7 @@ define void @v_shuffle_v4f32_v2f32__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -2166,8 +2145,7 @@ define void @v_shuffle_v4f32_v2f32__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -2246,16 +2224,15 @@ define void @v_shuffle_v4f32_v2f32__3_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:3]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2263,16 +2240,15 @@ define void @v_shuffle_v4f32_v2f32__3_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:3]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x float> asm "; def $0", "=v"()
@@ -2516,15 +2492,13 @@ define void @v_shuffle_v4f32_v2f32__3_3_1_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v5
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2534,15 +2508,13 @@ define void @v_shuffle_v4f32_v2f32__3_3_1_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x float> asm "; def $0", "=v"()
@@ -2826,16 +2798,15 @@ define void @v_shuffle_v4f32_v2f32__3_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:3]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2843,16 +2814,15 @@ define void @v_shuffle_v4f32_v2f32__3_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:3]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x float> asm "; def $0", "=v"()
@@ -2938,8 +2908,7 @@ define void @v_shuffle_v4f32_v2f32__3_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -2952,8 +2921,7 @@ define void @v_shuffle_v4f32_v2f32__3_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -3639,9 +3607,9 @@ define void @s_shuffle_v4f32_v2f32__3_3_3_0() {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def s[4:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: s_mov_b32 s10, s9
; GFX90A-NEXT: s_mov_b32 s11, s4
+; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; use s[8:11]
; GFX90A-NEXT: ;;#ASMEND
@@ -3656,9 +3624,9 @@ define void @s_shuffle_v4f32_v2f32__3_3_3_0() {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def s[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: s_mov_b32 s10, s9
; GFX940-NEXT: s_mov_b32 s11, s0
+; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; use s[8:11]
; GFX940-NEXT: ;;#ASMEND
@@ -3709,35 +3677,19 @@ define void @s_shuffle_v4f32_v2f32__3_3_3_2() {
; GFX900-NEXT: ;;#ASMEND
; GFX900-NEXT: s_setpc_b64 s[30:31]
;
-; GFX90A-LABEL: s_shuffle_v4f32_v2f32__3_3_3_2:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def s[4:5]
-; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s5
-; GFX90A-NEXT: s_mov_b32 s9, s5
-; GFX90A-NEXT: s_mov_b32 s10, s5
-; GFX90A-NEXT: s_mov_b32 s11, s4
-; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; use s[8:11]
-; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: s_shuffle_v4f32_v2f32__3_3_3_2:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def s[0:1]
-; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s1
-; GFX940-NEXT: s_mov_b32 s9, s1
-; GFX940-NEXT: s_mov_b32 s10, s1
-; GFX940-NEXT: s_mov_b32 s11, s0
-; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; use s[8:11]
-; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_setpc_b64 s[30:31]
+; GFX90APLUS-LABEL: s_shuffle_v4f32_v2f32__3_3_3_2:
+; GFX90APLUS: ; %bb.0:
+; GFX90APLUS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90APLUS-NEXT: ;;#ASMSTART
+; GFX90APLUS-NEXT: ; def s[8:9]
+; GFX90APLUS-NEXT: ;;#ASMEND
+; GFX90APLUS-NEXT: s_mov_b32 s10, s9
+; GFX90APLUS-NEXT: s_mov_b32 s11, s8
+; GFX90APLUS-NEXT: s_mov_b32 s8, s9
+; GFX90APLUS-NEXT: ;;#ASMSTART
+; GFX90APLUS-NEXT: ; use s[8:11]
+; GFX90APLUS-NEXT: ;;#ASMEND
+; GFX90APLUS-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x float> asm "; def $0", "=s"()
%vec1 = call <2 x float> asm "; def $0", "=s"()
%shuf = shufflevector <2 x float> %vec0, <2 x float> %vec1, <4 x i32> <i32 3, i32 3, i32 3, i32 2>
@@ -4178,9 +4130,9 @@ define void @s_shuffle_v4f32_v2f32__3_3_1_0() {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def s[4:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: s_mov_b32 s10, s5
; GFX90A-NEXT: s_mov_b32 s11, s4
+; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; use s[8:11]
; GFX90A-NEXT: ;;#ASMEND
@@ -4195,9 +4147,9 @@ define void @s_shuffle_v4f32_v2f32__3_3_1_0() {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def s[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: s_mov_b32 s10, s1
; GFX940-NEXT: s_mov_b32 s11, s0
+; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; use s[8:11]
; GFX940-NEXT: ;;#ASMEND
@@ -5081,15 +5033,14 @@ define void @s_shuffle_v4f32_v2f32__3_3_1_2() {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def s[4:5]
+; GFX90A-NEXT: ; def s[8:9]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def s[6:7]
+; GFX90A-NEXT: ; def s[4:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s7
-; GFX90A-NEXT: s_mov_b32 s9, s7
; GFX90A-NEXT: s_mov_b32 s10, s5
-; GFX90A-NEXT: s_mov_b32 s11, s6
+; GFX90A-NEXT: s_mov_b32 s11, s8
+; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; use s[8:11]
; GFX90A-NEXT: ;;#ASMEND
@@ -5099,15 +5050,14 @@ define void @s_shuffle_v4f32_v2f32__3_3_1_2() {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def s[0:1]
+; GFX940-NEXT: ; def s[8:9]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def s[2:3]
+; GFX940-NEXT: ; def s[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s3
-; GFX940-NEXT: s_mov_b32 s9, s3
; GFX940-NEXT: s_mov_b32 s10, s1
-; GFX940-NEXT: s_mov_b32 s11, s2
+; GFX940-NEXT: s_mov_b32 s11, s8
+; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; use s[8:11]
; GFX940-NEXT: ;;#ASMEND
@@ -5482,5 +5432,3 @@ define void @s_shuffle_v4f32_v2f32__3_3_2_3() {
call void asm sideeffect "; use $0", "{s[8:11]}"(<4 x float> %shuf)
ret void
}
-;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; GFX90APLUS: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v3f32.ll b/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v3f32.ll
index c797ef338dee96..31c458f5338cb9 100644
--- a/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v3f32.ll
+++ b/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v3f32.ll
@@ -1280,8 +1280,7 @@ define void @v_shuffle_v4f32_v3f32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
@@ -1296,8 +1295,7 @@ define void @v_shuffle_v4f32_v3f32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
@@ -1435,8 +1433,7 @@ define void @v_shuffle_v4f32_v3f32__4_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
@@ -1453,8 +1450,8 @@ define void @v_shuffle_v4f32_v3f32__4_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
@@ -1964,15 +1961,14 @@ define void @v_shuffle_v4f32_v3f32__5_5_1_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v2
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1982,15 +1978,15 @@ define void @v_shuffle_v4f32_v3f32__5_5_1_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v2
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x float> asm "; def $0", "=v"()
@@ -2142,17 +2138,16 @@ define void @v_shuffle_v4f32_v3f32__5_5_4_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v9, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[6:8]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v8
-; GFX90A-NEXT: v_mov_b32_e32 v1, v8
-; GFX90A-NEXT: v_mov_b32_e32 v2, v7
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v9, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2160,17 +2155,17 @@ define void @v_shuffle_v4f32_v3f32__5_5_4_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v9, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[6:8]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: v_mov_b32_e32 v0, v8
-; GFX940-NEXT: v_mov_b32_e32 v1, v8
-; GFX940-NEXT: v_mov_b32_e32 v2, v7
-; GFX940-NEXT: global_store_dwordx4 v9, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x float> asm "; def $0", "=v"()
@@ -3277,9 +3272,9 @@ define void @v_shuffle_v4f32_v3f32__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -3292,8 +3287,8 @@ define void @v_shuffle_v4f32_v3f32__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -3421,12 +3416,12 @@ define void @v_shuffle_v4f32_v3f32__4_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -3439,12 +3434,12 @@ define void @v_shuffle_v4f32_v3f32__4_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -3991,17 +3986,17 @@ define void @v_shuffle_v4f32_v3f32__5_5_1_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
-; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v5, 0
; GFX90A-NEXT: v_mov_b32_e32 v0, v2
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v7, 0
+; GFX90A-NEXT: ;;#ASMSTART
+; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v6
+; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4009,18 +4004,17 @@ define void @v_shuffle_v4f32_v3f32__5_5_1_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
-; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v5, 0
-; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_mov_b32_e32 v7, 0
; GFX940-NEXT: v_mov_b32_e32 v0, v2
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: ;;#ASMSTART
+; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x float> asm "; def $0", "=v"()
@@ -4112,17 +4106,17 @@ define void @v_shuffle_v4f32_v3f32__5_5_4_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[6:8]
+; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v8
-; GFX90A-NEXT: v_mov_b32_e32 v1, v8
-; GFX90A-NEXT: v_mov_b32_e32 v2, v7
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v6
+; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4130,17 +4124,18 @@ define void @v_shuffle_v4f32_v3f32__5_5_4_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v5, 0
+; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[6:8]
+; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: v_mov_b32_e32 v0, v8
-; GFX940-NEXT: v_mov_b32_e32 v1, v8
-; GFX940-NEXT: v_mov_b32_e32 v2, v7
-; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x float> asm "; def $0", "=v"()
@@ -4317,8 +4312,7 @@ define void @v_shuffle_v4f32_v3f32__4_3_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
@@ -4333,8 +4327,7 @@ define void @v_shuffle_v4f32_v3f32__4_3_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
@@ -4858,17 +4851,16 @@ define void @v_shuffle_v4f32_v3f32__5_5_1_3(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v6
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4876,17 +4868,17 @@ define void @v_shuffle_v4f32_v3f32__5_5_1_3(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v6
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x float> asm "; def $0", "=v"()
@@ -4972,31 +4964,28 @@ define void @v_shuffle_v4f32_v3f32__5_5_4_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4f32_v3f32__5_5_4_3:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v6
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4f32_v3f32__5_5_4_3:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v6
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x float> asm "; def $0", "=v"()
@@ -6094,9 +6083,9 @@ define void @v_shuffle_v4f32_v3f32__1_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v5, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6113,8 +6102,8 @@ define void @v_shuffle_v4f32_v3f32__1_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -6252,9 +6241,9 @@ define void @v_shuffle_v4f32_v3f32__4_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6267,8 +6256,8 @@ define void @v_shuffle_v4f32_v3f32__4_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -6727,17 +6716,16 @@ define void @v_shuffle_v4f32_v3f32__5_5_1_5(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v6
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v6
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -6745,17 +6733,17 @@ define void @v_shuffle_v4f32_v3f32__5_5_1_5(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v6
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
-; GFX940-NEXT: v_mov_b32_e32 v3, v6
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x float> asm "; def $0", "=v"()
@@ -6893,11 +6881,10 @@ define void @v_shuffle_v4f32_v3f32__5_5_4_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v5, 0
; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v1, v4
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -6910,9 +6897,8 @@ define void @v_shuffle_v4f32_v3f32__5_5_4_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v1, v4
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v4f32.ll b/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v4f32.ll
index 40d6ad90ab34b2..e3427cd35c683a 100644
--- a/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v4f32.ll
+++ b/llvm/test/CodeGen/AMDGPU/shufflevector.v4f32.v4f32.ll
@@ -343,13 +343,12 @@ define void @v_shuffle_v4f32_v4f32__7_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
-; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -364,9 +363,8 @@ define void @v_shuffle_v4f32_v4f32__7_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
-; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -456,8 +454,7 @@ define void @v_shuffle_v4f32_v4f32__7_2_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -472,8 +469,8 @@ define void @v_shuffle_v4f32_v4f32__7_2_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -560,9 +557,8 @@ define void @v_shuffle_v4f32_v4f32__7_4_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
-; GFX90A-NEXT: global_store_dwordx4 v4, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -573,9 +569,8 @@ define void @v_shuffle_v4f32_v4f32__7_4_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
-; GFX940-NEXT: global_store_dwordx4 v4, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -649,8 +644,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -662,8 +656,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -1169,11 +1162,10 @@ define void @v_shuffle_v4f32_v4f32__7_7_7_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v5
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1188,11 +1180,10 @@ define void @v_shuffle_v4f32_v4f32__7_7_7_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v5
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -1287,17 +1278,16 @@ define void @v_shuffle_v4f32_v4f32__7_7_7_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v7
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1305,17 +1295,17 @@ define void @v_shuffle_v4f32_v4f32__7_7_7_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v7
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -1402,31 +1392,26 @@ define void @v_shuffle_v4f32_v4f32__7_7_7_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4f32_v4f32__7_7_7_4:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4f32_v4f32__7_7_7_4:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -1508,31 +1493,26 @@ define void @v_shuffle_v4f32_v4f32__7_7_7_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4f32_v4f32__7_7_7_6:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4f32_v4f32__7_7_7_6:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -1709,8 +1689,7 @@ define void @v_shuffle_v4f32_v4f32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -1724,8 +1703,7 @@ define void @v_shuffle_v4f32_v4f32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -1809,8 +1787,7 @@ define void @v_shuffle_v4f32_v4f32__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -1824,8 +1801,7 @@ define void @v_shuffle_v4f32_v4f32__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -1914,8 +1890,7 @@ define void @v_shuffle_v4f32_v4f32__5_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -1933,8 +1908,7 @@ define void @v_shuffle_v4f32_v4f32__5_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -2038,8 +2012,7 @@ define void @v_shuffle_v4f32_v4f32__7_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -2057,8 +2030,7 @@ define void @v_shuffle_v4f32_v4f32__7_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -2214,17 +2186,16 @@ define void @v_shuffle_v4f32_v4f32__7_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v0
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v8, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v0
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: global_store_dwordx4 v8, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2232,17 +2203,16 @@ define void @v_shuffle_v4f32_v4f32__7_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v0
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: global_store_dwordx4 v8, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -2339,12 +2309,11 @@ define void @v_shuffle_v4f32_v4f32__7_4_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v0
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v0
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2354,15 +2323,15 @@ define void @v_shuffle_v4f32_v4f32__7_4_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -2460,8 +2429,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -2479,8 +2447,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -2643,11 +2610,10 @@ define void @v_shuffle_v4f32_v4f32__7_7_1_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v5
-; GFX90A-NEXT: v_mov_b32_e32 v4, v1
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2662,11 +2628,10 @@ define void @v_shuffle_v4f32_v4f32__7_7_1_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v5
-; GFX940-NEXT: v_mov_b32_e32 v4, v1
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -2759,17 +2724,16 @@ define void @v_shuffle_v4f32_v4f32__7_7_3_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2777,17 +2741,16 @@ define void @v_shuffle_v4f32_v4f32__7_7_3_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -2883,11 +2846,11 @@ define void @v_shuffle_v4f32_v4f32__7_7_5_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2897,14 +2860,15 @@ define void @v_shuffle_v4f32_v4f32__7_7_5_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -3517,8 +3481,7 @@ define void @v_shuffle_v4f32_v4f32__7_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v1
; GFX90A-NEXT: v_mov_b32_e32 v5, v1
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -3536,8 +3499,7 @@ define void @v_shuffle_v4f32_v4f32__7_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v1
; GFX940-NEXT: v_mov_b32_e32 v5, v1
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -3573,17 +3535,16 @@ define void @v_shuffle_v4f32_v4f32__7_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v1
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v8, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v1
+; GFX90A-NEXT: v_mov_b32_e32 v5, v1
+; GFX90A-NEXT: global_store_dwordx4 v8, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -3591,17 +3552,16 @@ define void @v_shuffle_v4f32_v4f32__7_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v1
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v1
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: global_store_dwordx4 v8, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -3697,12 +3657,11 @@ define void @v_shuffle_v4f32_v4f32__7_4_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v1
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v1
+; GFX90A-NEXT: v_mov_b32_e32 v5, v1
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -3712,15 +3671,15 @@ define void @v_shuffle_v4f32_v4f32__7_4_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v1
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v1
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -3818,8 +3777,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v1
; GFX90A-NEXT: v_mov_b32_e32 v5, v1
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -3837,8 +3795,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v1
; GFX940-NEXT: v_mov_b32_e32 v5, v1
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -4444,8 +4401,7 @@ define void @v_shuffle_v4f32_v4f32__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4458,8 +4414,7 @@ define void @v_shuffle_v4f32_v4f32__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4540,8 +4495,7 @@ define void @v_shuffle_v4f32_v4f32__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4554,8 +4508,7 @@ define void @v_shuffle_v4f32_v4f32__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4639,8 +4592,7 @@ define void @v_shuffle_v4f32_v4f32__5_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4656,8 +4608,8 @@ define void @v_shuffle_v4f32_v4f32__5_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4755,8 +4707,7 @@ define void @v_shuffle_v4f32_v4f32__7_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4772,8 +4723,8 @@ define void @v_shuffle_v4f32_v4f32__7_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4863,17 +4814,15 @@ define void @v_shuffle_v4f32_v4f32__7_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v2
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4881,17 +4830,15 @@ define void @v_shuffle_v4f32_v4f32__7_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -5042,8 +4989,7 @@ define void @v_shuffle_v4f32_v4f32__7_4_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -5060,8 +5006,7 @@ define void @v_shuffle_v4f32_v4f32__7_4_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -5158,8 +5103,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -5176,8 +5120,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -5388,17 +5331,16 @@ define void @v_shuffle_v4f32_v4f32__7_7_1_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -5406,17 +5348,16 @@ define void @v_shuffle_v4f32_v4f32__7_7_1_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -5449,17 +5390,16 @@ define void @v_shuffle_v4f32_v4f32__7_7_3_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -5467,17 +5407,16 @@ define void @v_shuffle_v4f32_v4f32__7_7_3_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -5568,16 +5507,16 @@ define void @v_shuffle_v4f32_v4f32__7_7_5_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v10, 0
+; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, v7
-; GFX90A-NEXT: v_mov_b32_e32 v8, v5
-; GFX90A-NEXT: v_mov_b32_e32 v9, v2
-; GFX90A-NEXT: global_store_dwordx4 v10, v[6:9], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -5585,16 +5524,17 @@ define void @v_shuffle_v4f32_v4f32__7_7_5_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v10, 0
+; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v7
-; GFX940-NEXT: v_mov_b32_e32 v8, v5
-; GFX940-NEXT: v_mov_b32_e32 v9, v2
-; GFX940-NEXT: global_store_dwordx4 v10, v[6:9], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -6193,17 +6133,15 @@ define void @v_shuffle_v4f32_v4f32__7_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -6211,17 +6149,15 @@ define void @v_shuffle_v4f32_v4f32__7_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -6314,8 +6250,7 @@ define void @v_shuffle_v4f32_v4f32__7_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6331,8 +6266,8 @@ define void @v_shuffle_v4f32_v4f32__7_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -6372,8 +6307,7 @@ define void @v_shuffle_v4f32_v4f32__7_4_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6390,8 +6324,7 @@ define void @v_shuffle_v4f32_v4f32__7_4_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -6488,8 +6421,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6506,8 +6438,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -7195,8 +7126,7 @@ define void @v_shuffle_v4f32_v4f32__5_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -7210,8 +7140,7 @@ define void @v_shuffle_v4f32_v4f32__5_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -7297,8 +7226,7 @@ define void @v_shuffle_v4f32_v4f32__7_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -7312,8 +7240,7 @@ define void @v_shuffle_v4f32_v4f32__7_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -7404,8 +7331,7 @@ define void @v_shuffle_v4f32_v4f32__7_0_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v6, v2
; GFX90A-NEXT: v_mov_b32_e32 v7, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
@@ -7423,8 +7349,7 @@ define void @v_shuffle_v4f32_v4f32__7_0_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v6, v2
; GFX940-NEXT: v_mov_b32_e32 v7, v2
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
@@ -7523,8 +7448,7 @@ define void @v_shuffle_v4f32_v4f32__7_2_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
@@ -7541,8 +7465,8 @@ define void @v_shuffle_v4f32_v4f32__7_2_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
@@ -7686,31 +7610,28 @@ define void @v_shuffle_v4f32_v4f32__7_6_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4f32_v4f32__7_6_4_4:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v0
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v0
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4f32_v4f32__7_6_4_4:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v0
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -7904,11 +7825,11 @@ define void @v_shuffle_v4f32_v4f32__7_7_1_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -7918,14 +7839,15 @@ define void @v_shuffle_v4f32_v4f32__7_7_1_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -8022,10 +7944,9 @@ define void @v_shuffle_v4f32_v4f32__7_7_3_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v7
; GFX90A-NEXT: v_mov_b32_e32 v1, v7
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -8040,10 +7961,10 @@ define void @v_shuffle_v4f32_v4f32__7_7_3_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v7
; GFX940-NEXT: v_mov_b32_e32 v1, v7
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -8076,9 +7997,8 @@ define void @v_shuffle_v4f32_v4f32__7_7_5_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v4, v1
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -8090,9 +8010,8 @@ define void @v_shuffle_v4f32_v4f32__7_7_5_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v4, v1
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -8703,12 +8622,10 @@ define void @v_shuffle_v4f32_v4f32__7_0_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -8718,16 +8635,14 @@ define void @v_shuffle_v4f32_v4f32__7_0_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -8822,8 +8737,7 @@ define void @v_shuffle_v4f32_v4f32__7_2_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v5
; GFX90A-NEXT: v_mov_b32_e32 v3, v5
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
@@ -8840,8 +8754,8 @@ define void @v_shuffle_v4f32_v4f32__7_2_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v5
; GFX940-NEXT: v_mov_b32_e32 v3, v5
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
@@ -8938,8 +8852,7 @@ define void @v_shuffle_v4f32_v4f32__7_4_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v1
; GFX90A-NEXT: v_mov_b32_e32 v5, v1
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -8953,8 +8866,7 @@ define void @v_shuffle_v4f32_v4f32__7_4_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v1
; GFX940-NEXT: v_mov_b32_e32 v5, v1
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -8985,31 +8897,28 @@ define void @v_shuffle_v4f32_v4f32__7_6_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4f32_v4f32__7_6_5_5:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v1
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v1
+; GFX90A-NEXT: v_mov_b32_e32 v5, v1
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4f32_v4f32__7_6_5_5:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v1
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v1
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -9584,8 +9493,7 @@ define void @v_shuffle_v4f32_v4f32__1_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v5, v4
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9602,8 +9510,7 @@ define void @v_shuffle_v4f32_v4f32__1_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v5, v4
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9701,8 +9608,7 @@ define void @v_shuffle_v4f32_v4f32__3_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v6
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v7, v6
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9718,8 +9624,8 @@ define void @v_shuffle_v4f32_v4f32__3_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v5, v6
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[6:7] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v7, v6
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9799,8 +9705,7 @@ define void @v_shuffle_v4f32_v4f32__5_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9813,8 +9718,7 @@ define void @v_shuffle_v4f32_v4f32__5_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9897,8 +9801,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9911,8 +9814,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9998,8 +9900,7 @@ define void @v_shuffle_v4f32_v4f32__7_0_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v5, v4
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -10016,8 +9917,7 @@ define void @v_shuffle_v4f32_v4f32__7_0_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v5, v4
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -10116,8 +10016,7 @@ define void @v_shuffle_v4f32_v4f32__7_2_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v7, v6
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -10133,8 +10032,8 @@ define void @v_shuffle_v4f32_v4f32__7_2_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v7, v6
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -10222,31 +10121,26 @@ define void @v_shuffle_v4f32_v4f32__7_4_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4f32_v4f32__7_4_6_6:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v2
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4f32_v4f32__7_4_6_6:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -10486,15 +10380,14 @@ define void @v_shuffle_v4f32_v4f32__7_7_1_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v10, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, v5
-; GFX90A-NEXT: v_mov_b32_e32 v7, v5
-; GFX90A-NEXT: v_mov_b32_e32 v8, v1
-; GFX90A-NEXT: v_mov_b32_e32 v9, v4
-; GFX90A-NEXT: global_store_dwordx4 v10, v[6:9], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -10504,15 +10397,15 @@ define void @v_shuffle_v4f32_v4f32__7_7_1_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v10, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, v1
-; GFX940-NEXT: v_mov_b32_e32 v6, v5
-; GFX940-NEXT: v_mov_b32_e32 v7, v5
-; GFX940-NEXT: v_mov_b32_e32 v9, v4
-; GFX940-NEXT: global_store_dwordx4 v10, v[6:9], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -10610,10 +10503,9 @@ define void @v_shuffle_v4f32_v4f32__7_7_3_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v7
; GFX90A-NEXT: v_mov_b32_e32 v1, v7
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v6
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -10628,10 +10520,10 @@ define void @v_shuffle_v4f32_v4f32__7_7_3_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[6:7] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v7
; GFX940-NEXT: v_mov_b32_e32 v1, v7
-; GFX940-NEXT: v_mov_b32_e32 v3, v6
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -10714,31 +10606,26 @@ define void @v_shuffle_v4f32_v4f32__7_7_5_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4f32_v4f32__7_7_5_6:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4f32_v4f32__7_7_5_6:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -11246,8 +11133,7 @@ define void @v_shuffle_v4f32_v4f32__7_0_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v5
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -11264,8 +11150,7 @@ define void @v_shuffle_v4f32_v4f32__7_0_7_7(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v5
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -11364,8 +11249,7 @@ define void @v_shuffle_v4f32_v4f32__7_2_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v6, v7
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -11381,8 +11265,8 @@ define void @v_shuffle_v4f32_v4f32__7_2_7_7(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v6, v7
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -11471,31 +11355,26 @@ define void @v_shuffle_v4f32_v4f32__7_4_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4f32_v4f32__7_4_7_7:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4f32_v4f32__7_4_7_7:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x float> asm "; def $0", "=v"()
@@ -11573,8 +11452,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -11587,8 +11465,7 @@ define void @v_shuffle_v4f32_v4f32__7_6_7_7(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v2i32.ll b/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v2i32.ll
index 86e8e2ed267dc6..e2cfbe0e9e1727 100644
--- a/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v2i32.ll
+++ b/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v2i32.ll
@@ -176,8 +176,7 @@ define void @v_shuffle_v4i32_v2i32__3_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -192,8 +191,8 @@ define void @v_shuffle_v4i32_v2i32__3_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -566,16 +565,15 @@ define void @v_shuffle_v4i32_v2i32__3_3_3_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:1]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -583,16 +581,15 @@ define void @v_shuffle_v4i32_v2i32__3_3_3_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:1]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x i32> asm "; def $0", "=v"()
@@ -675,31 +672,26 @@ define void @v_shuffle_v4i32_v2i32__3_3_3_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4i32_v2i32__3_3_3_2:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v5
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4i32_v2i32__3_3_3_2:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v5
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x i32> asm "; def $0", "=v"()
@@ -876,8 +868,7 @@ define void @v_shuffle_v4i32_v2i32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -890,8 +881,7 @@ define void @v_shuffle_v4i32_v2i32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -978,8 +968,7 @@ define void @v_shuffle_v4i32_v2i32__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -995,8 +984,7 @@ define void @v_shuffle_v4i32_v2i32__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -1144,14 +1132,13 @@ define void @v_shuffle_v4i32_v2i32__3_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1161,14 +1148,13 @@ define void @v_shuffle_v4i32_v2i32__3_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x i32> asm "; def $0", "=v"()
@@ -1309,16 +1295,15 @@ define void @v_shuffle_v4i32_v2i32__3_3_1_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:1]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1326,16 +1311,15 @@ define void @v_shuffle_v4i32_v2i32__3_3_1_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:1]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x i32> asm "; def $0", "=v"()
@@ -1729,8 +1713,7 @@ define void @v_shuffle_v4i32_v2i32__3_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -1746,8 +1729,7 @@ define void @v_shuffle_v4i32_v2i32__3_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -1782,14 +1764,13 @@ define void @v_shuffle_v4i32_v2i32__3_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1799,14 +1780,13 @@ define void @v_shuffle_v4i32_v2i32__3_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x i32> asm "; def $0", "=v"()
@@ -2158,8 +2138,7 @@ define void @v_shuffle_v4i32_v2i32__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -2172,8 +2151,7 @@ define void @v_shuffle_v4i32_v2i32__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -2252,16 +2230,15 @@ define void @v_shuffle_v4i32_v2i32__3_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:3]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2269,16 +2246,15 @@ define void @v_shuffle_v4i32_v2i32__3_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:3]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x i32> asm "; def $0", "=v"()
@@ -2522,15 +2498,13 @@ define void @v_shuffle_v4i32_v2i32__3_3_1_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v5
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2540,15 +2514,13 @@ define void @v_shuffle_v4i32_v2i32__3_3_1_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x i32> asm "; def $0", "=v"()
@@ -2832,16 +2804,15 @@ define void @v_shuffle_v4i32_v2i32__3_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:3]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2849,16 +2820,15 @@ define void @v_shuffle_v4i32_v2i32__3_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:3]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x i32> asm "; def $0", "=v"()
@@ -2944,8 +2914,7 @@ define void @v_shuffle_v4i32_v2i32__3_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -2958,8 +2927,7 @@ define void @v_shuffle_v4i32_v2i32__3_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -3645,9 +3613,9 @@ define void @s_shuffle_v4i32_v2i32__3_3_3_0() {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def s[4:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: s_mov_b32 s10, s9
; GFX90A-NEXT: s_mov_b32 s11, s4
+; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; use s[8:11]
; GFX90A-NEXT: ;;#ASMEND
@@ -3662,9 +3630,9 @@ define void @s_shuffle_v4i32_v2i32__3_3_3_0() {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def s[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: s_mov_b32 s10, s9
; GFX940-NEXT: s_mov_b32 s11, s0
+; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; use s[8:11]
; GFX940-NEXT: ;;#ASMEND
@@ -3715,35 +3683,19 @@ define void @s_shuffle_v4i32_v2i32__3_3_3_2() {
; GFX900-NEXT: ;;#ASMEND
; GFX900-NEXT: s_setpc_b64 s[30:31]
;
-; GFX90A-LABEL: s_shuffle_v4i32_v2i32__3_3_3_2:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def s[4:5]
-; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s5
-; GFX90A-NEXT: s_mov_b32 s9, s5
-; GFX90A-NEXT: s_mov_b32 s10, s5
-; GFX90A-NEXT: s_mov_b32 s11, s4
-; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; use s[8:11]
-; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: s_shuffle_v4i32_v2i32__3_3_3_2:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def s[0:1]
-; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s1
-; GFX940-NEXT: s_mov_b32 s9, s1
-; GFX940-NEXT: s_mov_b32 s10, s1
-; GFX940-NEXT: s_mov_b32 s11, s0
-; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; use s[8:11]
-; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_setpc_b64 s[30:31]
+; GFX90APLUS-LABEL: s_shuffle_v4i32_v2i32__3_3_3_2:
+; GFX90APLUS: ; %bb.0:
+; GFX90APLUS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90APLUS-NEXT: ;;#ASMSTART
+; GFX90APLUS-NEXT: ; def s[8:9]
+; GFX90APLUS-NEXT: ;;#ASMEND
+; GFX90APLUS-NEXT: s_mov_b32 s10, s9
+; GFX90APLUS-NEXT: s_mov_b32 s11, s8
+; GFX90APLUS-NEXT: s_mov_b32 s8, s9
+; GFX90APLUS-NEXT: ;;#ASMSTART
+; GFX90APLUS-NEXT: ; use s[8:11]
+; GFX90APLUS-NEXT: ;;#ASMEND
+; GFX90APLUS-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x i32> asm "; def $0", "=s"()
%vec1 = call <2 x i32> asm "; def $0", "=s"()
%shuf = shufflevector <2 x i32> %vec0, <2 x i32> %vec1, <4 x i32> <i32 3, i32 3, i32 3, i32 2>
@@ -4186,9 +4138,9 @@ define void @s_shuffle_v4i32_v2i32__3_3_1_0() {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def s[4:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: s_mov_b32 s10, s5
; GFX90A-NEXT: s_mov_b32 s11, s4
+; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; use s[8:11]
; GFX90A-NEXT: ;;#ASMEND
@@ -4203,9 +4155,9 @@ define void @s_shuffle_v4i32_v2i32__3_3_1_0() {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def s[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: s_mov_b32 s10, s1
; GFX940-NEXT: s_mov_b32 s11, s0
+; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; use s[8:11]
; GFX940-NEXT: ;;#ASMEND
@@ -5089,15 +5041,14 @@ define void @s_shuffle_v4i32_v2i32__3_3_1_2() {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def s[4:5]
+; GFX90A-NEXT: ; def s[8:9]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def s[6:7]
+; GFX90A-NEXT: ; def s[4:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s7
-; GFX90A-NEXT: s_mov_b32 s9, s7
; GFX90A-NEXT: s_mov_b32 s10, s5
-; GFX90A-NEXT: s_mov_b32 s11, s6
+; GFX90A-NEXT: s_mov_b32 s11, s8
+; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; use s[8:11]
; GFX90A-NEXT: ;;#ASMEND
@@ -5107,15 +5058,14 @@ define void @s_shuffle_v4i32_v2i32__3_3_1_2() {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def s[0:1]
+; GFX940-NEXT: ; def s[8:9]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def s[2:3]
+; GFX940-NEXT: ; def s[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s3
-; GFX940-NEXT: s_mov_b32 s9, s3
; GFX940-NEXT: s_mov_b32 s10, s1
-; GFX940-NEXT: s_mov_b32 s11, s2
+; GFX940-NEXT: s_mov_b32 s11, s8
+; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; use s[8:11]
; GFX940-NEXT: ;;#ASMEND
@@ -5490,5 +5440,3 @@ define void @s_shuffle_v4i32_v2i32__3_3_2_3() {
call void asm sideeffect "; use $0", "{s[8:11]}"(<4 x i32> %shuf)
ret void
}
-;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; GFX90APLUS: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v3i32.ll b/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v3i32.ll
index 770c1c19087b4e..1b003a7c5d9bc2 100644
--- a/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v3i32.ll
+++ b/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v3i32.ll
@@ -1280,8 +1280,7 @@ define void @v_shuffle_v4i32_v3i32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
@@ -1296,8 +1295,7 @@ define void @v_shuffle_v4i32_v3i32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
@@ -1435,8 +1433,7 @@ define void @v_shuffle_v4i32_v3i32__4_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
@@ -1453,8 +1450,8 @@ define void @v_shuffle_v4i32_v3i32__4_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
@@ -1964,15 +1961,14 @@ define void @v_shuffle_v4i32_v3i32__5_5_1_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v2
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1982,15 +1978,15 @@ define void @v_shuffle_v4i32_v3i32__5_5_1_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v2
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x i32> asm "; def $0", "=v"()
@@ -2142,17 +2138,16 @@ define void @v_shuffle_v4i32_v3i32__5_5_4_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v9, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[6:8]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v8
-; GFX90A-NEXT: v_mov_b32_e32 v1, v8
-; GFX90A-NEXT: v_mov_b32_e32 v2, v7
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v9, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2160,17 +2155,17 @@ define void @v_shuffle_v4i32_v3i32__5_5_4_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v9, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[6:8]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: v_mov_b32_e32 v0, v8
-; GFX940-NEXT: v_mov_b32_e32 v1, v8
-; GFX940-NEXT: v_mov_b32_e32 v2, v7
-; GFX940-NEXT: global_store_dwordx4 v9, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x i32> asm "; def $0", "=v"()
@@ -3277,9 +3272,9 @@ define void @v_shuffle_v4i32_v3i32__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -3292,8 +3287,8 @@ define void @v_shuffle_v4i32_v3i32__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -3421,12 +3416,12 @@ define void @v_shuffle_v4i32_v3i32__4_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -3439,12 +3434,12 @@ define void @v_shuffle_v4i32_v3i32__4_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -3991,17 +3986,17 @@ define void @v_shuffle_v4i32_v3i32__5_5_1_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
-; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v5, 0
; GFX90A-NEXT: v_mov_b32_e32 v0, v2
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v7, 0
+; GFX90A-NEXT: ;;#ASMSTART
+; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v6
+; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4009,18 +4004,17 @@ define void @v_shuffle_v4i32_v3i32__5_5_1_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
-; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v5, 0
-; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_mov_b32_e32 v7, 0
; GFX940-NEXT: v_mov_b32_e32 v0, v2
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: ;;#ASMSTART
+; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x i32> asm "; def $0", "=v"()
@@ -4112,17 +4106,17 @@ define void @v_shuffle_v4i32_v3i32__5_5_4_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[6:8]
+; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v8
-; GFX90A-NEXT: v_mov_b32_e32 v1, v8
-; GFX90A-NEXT: v_mov_b32_e32 v2, v7
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v6
+; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4130,17 +4124,18 @@ define void @v_shuffle_v4i32_v3i32__5_5_4_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v5, 0
+; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[6:8]
+; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: v_mov_b32_e32 v0, v8
-; GFX940-NEXT: v_mov_b32_e32 v1, v8
-; GFX940-NEXT: v_mov_b32_e32 v2, v7
-; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x i32> asm "; def $0", "=v"()
@@ -4317,8 +4312,7 @@ define void @v_shuffle_v4i32_v3i32__4_3_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
@@ -4333,8 +4327,7 @@ define void @v_shuffle_v4i32_v3i32__4_3_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
@@ -4858,17 +4851,16 @@ define void @v_shuffle_v4i32_v3i32__5_5_1_3(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v6
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4876,17 +4868,17 @@ define void @v_shuffle_v4i32_v3i32__5_5_1_3(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v6
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x i32> asm "; def $0", "=v"()
@@ -4972,31 +4964,28 @@ define void @v_shuffle_v4i32_v3i32__5_5_4_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4i32_v3i32__5_5_4_3:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v6
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4i32_v3i32__5_5_4_3:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v6
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x i32> asm "; def $0", "=v"()
@@ -6094,9 +6083,9 @@ define void @v_shuffle_v4i32_v3i32__1_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v5, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6113,8 +6102,8 @@ define void @v_shuffle_v4i32_v3i32__1_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -6252,9 +6241,9 @@ define void @v_shuffle_v4i32_v3i32__4_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6267,8 +6256,8 @@ define void @v_shuffle_v4i32_v3i32__4_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -6727,17 +6716,16 @@ define void @v_shuffle_v4i32_v3i32__5_5_1_5(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v6
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v6
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -6745,17 +6733,17 @@ define void @v_shuffle_v4i32_v3i32__5_5_1_5(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v6
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
-; GFX940-NEXT: v_mov_b32_e32 v3, v6
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x i32> asm "; def $0", "=v"()
@@ -6893,11 +6881,10 @@ define void @v_shuffle_v4i32_v3i32__5_5_4_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v5, 0
; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v1, v4
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -6910,9 +6897,8 @@ define void @v_shuffle_v4i32_v3i32__5_5_4_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v1, v4
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v4i32.ll b/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v4i32.ll
index 3d1bb1da92e09b..47ad1c4bedb8b1 100644
--- a/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v4i32.ll
+++ b/llvm/test/CodeGen/AMDGPU/shufflevector.v4i32.v4i32.ll
@@ -343,13 +343,12 @@ define void @v_shuffle_v4i32_v4i32__7_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
-; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -364,9 +363,8 @@ define void @v_shuffle_v4i32_v4i32__7_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
-; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -456,8 +454,7 @@ define void @v_shuffle_v4i32_v4i32__7_2_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -472,8 +469,8 @@ define void @v_shuffle_v4i32_v4i32__7_2_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -560,9 +557,8 @@ define void @v_shuffle_v4i32_v4i32__7_4_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
-; GFX90A-NEXT: global_store_dwordx4 v4, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -573,9 +569,8 @@ define void @v_shuffle_v4i32_v4i32__7_4_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
-; GFX940-NEXT: global_store_dwordx4 v4, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -649,8 +644,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -662,8 +656,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -1169,11 +1162,10 @@ define void @v_shuffle_v4i32_v4i32__7_7_7_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v5
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1188,11 +1180,10 @@ define void @v_shuffle_v4i32_v4i32__7_7_7_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v5
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -1287,17 +1278,16 @@ define void @v_shuffle_v4i32_v4i32__7_7_7_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v7
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1305,17 +1295,17 @@ define void @v_shuffle_v4i32_v4i32__7_7_7_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v7
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -1402,31 +1392,26 @@ define void @v_shuffle_v4i32_v4i32__7_7_7_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4i32_v4i32__7_7_7_4:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4i32_v4i32__7_7_7_4:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -1508,31 +1493,26 @@ define void @v_shuffle_v4i32_v4i32__7_7_7_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4i32_v4i32__7_7_7_6:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4i32_v4i32__7_7_7_6:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -1709,8 +1689,7 @@ define void @v_shuffle_v4i32_v4i32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -1724,8 +1703,7 @@ define void @v_shuffle_v4i32_v4i32__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -1809,8 +1787,7 @@ define void @v_shuffle_v4i32_v4i32__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -1824,8 +1801,7 @@ define void @v_shuffle_v4i32_v4i32__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -1914,8 +1890,7 @@ define void @v_shuffle_v4i32_v4i32__5_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -1933,8 +1908,7 @@ define void @v_shuffle_v4i32_v4i32__5_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -2038,8 +2012,7 @@ define void @v_shuffle_v4i32_v4i32__7_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -2057,8 +2030,7 @@ define void @v_shuffle_v4i32_v4i32__7_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -2214,17 +2186,16 @@ define void @v_shuffle_v4i32_v4i32__7_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v0
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v8, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v0
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: global_store_dwordx4 v8, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2232,17 +2203,16 @@ define void @v_shuffle_v4i32_v4i32__7_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v0
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: global_store_dwordx4 v8, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -2339,12 +2309,11 @@ define void @v_shuffle_v4i32_v4i32__7_4_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v0
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v0
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2354,15 +2323,15 @@ define void @v_shuffle_v4i32_v4i32__7_4_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -2460,8 +2429,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -2479,8 +2447,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -2643,11 +2610,10 @@ define void @v_shuffle_v4i32_v4i32__7_7_1_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v5
-; GFX90A-NEXT: v_mov_b32_e32 v4, v1
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2662,11 +2628,10 @@ define void @v_shuffle_v4i32_v4i32__7_7_1_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v5
-; GFX940-NEXT: v_mov_b32_e32 v4, v1
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -2759,17 +2724,16 @@ define void @v_shuffle_v4i32_v4i32__7_7_3_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2777,17 +2741,16 @@ define void @v_shuffle_v4i32_v4i32__7_7_3_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -2883,11 +2846,11 @@ define void @v_shuffle_v4i32_v4i32__7_7_5_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2897,14 +2860,15 @@ define void @v_shuffle_v4i32_v4i32__7_7_5_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -3517,8 +3481,7 @@ define void @v_shuffle_v4i32_v4i32__7_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v1
; GFX90A-NEXT: v_mov_b32_e32 v5, v1
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -3536,8 +3499,7 @@ define void @v_shuffle_v4i32_v4i32__7_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v1
; GFX940-NEXT: v_mov_b32_e32 v5, v1
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -3573,17 +3535,16 @@ define void @v_shuffle_v4i32_v4i32__7_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v1
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v8, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v1
+; GFX90A-NEXT: v_mov_b32_e32 v5, v1
+; GFX90A-NEXT: global_store_dwordx4 v8, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -3591,17 +3552,16 @@ define void @v_shuffle_v4i32_v4i32__7_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v1
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v1
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: global_store_dwordx4 v8, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -3697,12 +3657,11 @@ define void @v_shuffle_v4i32_v4i32__7_4_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v1
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v1
+; GFX90A-NEXT: v_mov_b32_e32 v5, v1
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -3712,15 +3671,15 @@ define void @v_shuffle_v4i32_v4i32__7_4_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v1
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v1
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -3818,8 +3777,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v1
; GFX90A-NEXT: v_mov_b32_e32 v5, v1
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -3837,8 +3795,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v1
; GFX940-NEXT: v_mov_b32_e32 v5, v1
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -4444,8 +4401,7 @@ define void @v_shuffle_v4i32_v4i32__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4458,8 +4414,7 @@ define void @v_shuffle_v4i32_v4i32__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4540,8 +4495,7 @@ define void @v_shuffle_v4i32_v4i32__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4554,8 +4508,7 @@ define void @v_shuffle_v4i32_v4i32__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4639,8 +4592,7 @@ define void @v_shuffle_v4i32_v4i32__5_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4656,8 +4608,8 @@ define void @v_shuffle_v4i32_v4i32__5_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4755,8 +4707,7 @@ define void @v_shuffle_v4i32_v4i32__7_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4772,8 +4723,8 @@ define void @v_shuffle_v4i32_v4i32__7_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4863,17 +4814,15 @@ define void @v_shuffle_v4i32_v4i32__7_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v2
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4881,17 +4830,15 @@ define void @v_shuffle_v4i32_v4i32__7_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -5042,8 +4989,7 @@ define void @v_shuffle_v4i32_v4i32__7_4_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -5060,8 +5006,7 @@ define void @v_shuffle_v4i32_v4i32__7_4_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -5158,8 +5103,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -5176,8 +5120,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -5388,17 +5331,16 @@ define void @v_shuffle_v4i32_v4i32__7_7_1_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -5406,17 +5348,16 @@ define void @v_shuffle_v4i32_v4i32__7_7_1_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -5449,17 +5390,16 @@ define void @v_shuffle_v4i32_v4i32__7_7_3_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -5467,17 +5407,16 @@ define void @v_shuffle_v4i32_v4i32__7_7_3_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -5568,16 +5507,16 @@ define void @v_shuffle_v4i32_v4i32__7_7_5_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v10, 0
+; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, v7
-; GFX90A-NEXT: v_mov_b32_e32 v8, v5
-; GFX90A-NEXT: v_mov_b32_e32 v9, v2
-; GFX90A-NEXT: global_store_dwordx4 v10, v[6:9], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -5585,16 +5524,17 @@ define void @v_shuffle_v4i32_v4i32__7_7_5_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v10, 0
+; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v7
-; GFX940-NEXT: v_mov_b32_e32 v8, v5
-; GFX940-NEXT: v_mov_b32_e32 v9, v2
-; GFX940-NEXT: global_store_dwordx4 v10, v[6:9], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -6193,17 +6133,15 @@ define void @v_shuffle_v4i32_v4i32__7_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -6211,17 +6149,15 @@ define void @v_shuffle_v4i32_v4i32__7_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -6314,8 +6250,7 @@ define void @v_shuffle_v4i32_v4i32__7_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6331,8 +6266,8 @@ define void @v_shuffle_v4i32_v4i32__7_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -6372,8 +6307,7 @@ define void @v_shuffle_v4i32_v4i32__7_4_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6390,8 +6324,7 @@ define void @v_shuffle_v4i32_v4i32__7_4_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -6488,8 +6421,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6506,8 +6438,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -7195,8 +7126,7 @@ define void @v_shuffle_v4i32_v4i32__5_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -7210,8 +7140,7 @@ define void @v_shuffle_v4i32_v4i32__5_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -7297,8 +7226,7 @@ define void @v_shuffle_v4i32_v4i32__7_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -7312,8 +7240,7 @@ define void @v_shuffle_v4i32_v4i32__7_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -7404,8 +7331,7 @@ define void @v_shuffle_v4i32_v4i32__7_0_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v6, v2
; GFX90A-NEXT: v_mov_b32_e32 v7, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
@@ -7423,8 +7349,7 @@ define void @v_shuffle_v4i32_v4i32__7_0_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v6, v2
; GFX940-NEXT: v_mov_b32_e32 v7, v2
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
@@ -7523,8 +7448,7 @@ define void @v_shuffle_v4i32_v4i32__7_2_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
@@ -7541,8 +7465,8 @@ define void @v_shuffle_v4i32_v4i32__7_2_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
@@ -7686,31 +7610,28 @@ define void @v_shuffle_v4i32_v4i32__7_6_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4i32_v4i32__7_6_4_4:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v0
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v0
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4i32_v4i32__7_6_4_4:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v0
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -7904,11 +7825,11 @@ define void @v_shuffle_v4i32_v4i32__7_7_1_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -7918,14 +7839,15 @@ define void @v_shuffle_v4i32_v4i32__7_7_1_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -8022,10 +7944,9 @@ define void @v_shuffle_v4i32_v4i32__7_7_3_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v7
; GFX90A-NEXT: v_mov_b32_e32 v1, v7
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -8040,10 +7961,10 @@ define void @v_shuffle_v4i32_v4i32__7_7_3_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v7
; GFX940-NEXT: v_mov_b32_e32 v1, v7
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -8076,9 +7997,8 @@ define void @v_shuffle_v4i32_v4i32__7_7_5_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v4, v1
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -8090,9 +8010,8 @@ define void @v_shuffle_v4i32_v4i32__7_7_5_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v4, v1
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -8703,12 +8622,10 @@ define void @v_shuffle_v4i32_v4i32__7_0_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -8718,16 +8635,14 @@ define void @v_shuffle_v4i32_v4i32__7_0_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -8822,8 +8737,7 @@ define void @v_shuffle_v4i32_v4i32__7_2_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v5
; GFX90A-NEXT: v_mov_b32_e32 v3, v5
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
@@ -8840,8 +8754,8 @@ define void @v_shuffle_v4i32_v4i32__7_2_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v5
; GFX940-NEXT: v_mov_b32_e32 v3, v5
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
@@ -8938,8 +8852,7 @@ define void @v_shuffle_v4i32_v4i32__7_4_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v1
; GFX90A-NEXT: v_mov_b32_e32 v5, v1
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -8953,8 +8866,7 @@ define void @v_shuffle_v4i32_v4i32__7_4_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v1
; GFX940-NEXT: v_mov_b32_e32 v5, v1
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -8985,31 +8897,28 @@ define void @v_shuffle_v4i32_v4i32__7_6_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4i32_v4i32__7_6_5_5:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v1
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v1
+; GFX90A-NEXT: v_mov_b32_e32 v5, v1
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4i32_v4i32__7_6_5_5:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v1
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v1
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -9584,8 +9493,7 @@ define void @v_shuffle_v4i32_v4i32__1_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v5, v4
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9602,8 +9510,7 @@ define void @v_shuffle_v4i32_v4i32__1_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v5, v4
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9701,8 +9608,7 @@ define void @v_shuffle_v4i32_v4i32__3_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v6
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v7, v6
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9718,8 +9624,8 @@ define void @v_shuffle_v4i32_v4i32__3_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v5, v6
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[6:7] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v7, v6
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9799,8 +9705,7 @@ define void @v_shuffle_v4i32_v4i32__5_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9813,8 +9718,7 @@ define void @v_shuffle_v4i32_v4i32__5_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9897,8 +9801,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9911,8 +9814,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9998,8 +9900,7 @@ define void @v_shuffle_v4i32_v4i32__7_0_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v5, v4
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -10016,8 +9917,7 @@ define void @v_shuffle_v4i32_v4i32__7_0_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v5, v4
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -10116,8 +10016,7 @@ define void @v_shuffle_v4i32_v4i32__7_2_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v7, v6
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -10133,8 +10032,8 @@ define void @v_shuffle_v4i32_v4i32__7_2_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v7, v6
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -10222,31 +10121,26 @@ define void @v_shuffle_v4i32_v4i32__7_4_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4i32_v4i32__7_4_6_6:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v2
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4i32_v4i32__7_4_6_6:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -10486,15 +10380,14 @@ define void @v_shuffle_v4i32_v4i32__7_7_1_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v10, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, v5
-; GFX90A-NEXT: v_mov_b32_e32 v7, v5
-; GFX90A-NEXT: v_mov_b32_e32 v8, v1
-; GFX90A-NEXT: v_mov_b32_e32 v9, v4
-; GFX90A-NEXT: global_store_dwordx4 v10, v[6:9], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -10504,15 +10397,15 @@ define void @v_shuffle_v4i32_v4i32__7_7_1_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v10, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, v1
-; GFX940-NEXT: v_mov_b32_e32 v6, v5
-; GFX940-NEXT: v_mov_b32_e32 v7, v5
-; GFX940-NEXT: v_mov_b32_e32 v9, v4
-; GFX940-NEXT: global_store_dwordx4 v10, v[6:9], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -10610,10 +10503,9 @@ define void @v_shuffle_v4i32_v4i32__7_7_3_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v7
; GFX90A-NEXT: v_mov_b32_e32 v1, v7
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v6
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -10628,10 +10520,10 @@ define void @v_shuffle_v4i32_v4i32__7_7_3_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[6:7] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v7
; GFX940-NEXT: v_mov_b32_e32 v1, v7
-; GFX940-NEXT: v_mov_b32_e32 v3, v6
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -10714,31 +10606,26 @@ define void @v_shuffle_v4i32_v4i32__7_7_5_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4i32_v4i32__7_7_5_6:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4i32_v4i32__7_7_5_6:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -11246,8 +11133,7 @@ define void @v_shuffle_v4i32_v4i32__7_0_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v5
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -11264,8 +11150,7 @@ define void @v_shuffle_v4i32_v4i32__7_0_7_7(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v5
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -11364,8 +11249,7 @@ define void @v_shuffle_v4i32_v4i32__7_2_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v6, v7
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -11381,8 +11265,8 @@ define void @v_shuffle_v4i32_v4i32__7_2_7_7(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v6, v7
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -11471,31 +11355,26 @@ define void @v_shuffle_v4i32_v4i32__7_4_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4i32_v4i32__7_4_7_7:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4i32_v4i32__7_4_7_7:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x i32> asm "; def $0", "=v"()
@@ -11573,8 +11452,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -11587,8 +11465,7 @@ define void @v_shuffle_v4i32_v4i32__7_6_7_7(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v2p3.ll b/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v2p3.ll
index 2dd794404158c3..7a760e2da7dfca 100644
--- a/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v2p3.ll
+++ b/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v2p3.ll
@@ -176,8 +176,7 @@ define void @v_shuffle_v4p3_v2p3__3_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -192,8 +191,8 @@ define void @v_shuffle_v4p3_v2p3__3_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -566,16 +565,15 @@ define void @v_shuffle_v4p3_v2p3__3_3_3_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:1]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -583,16 +581,15 @@ define void @v_shuffle_v4p3_v2p3__3_3_3_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:1]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -675,31 +672,26 @@ define void @v_shuffle_v4p3_v2p3__3_3_3_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4p3_v2p3__3_3_3_2:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v5
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4p3_v2p3__3_3_3_2:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v5
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -876,8 +868,7 @@ define void @v_shuffle_v4p3_v2p3__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -890,8 +881,7 @@ define void @v_shuffle_v4p3_v2p3__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -978,8 +968,7 @@ define void @v_shuffle_v4p3_v2p3__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -995,8 +984,7 @@ define void @v_shuffle_v4p3_v2p3__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -1144,14 +1132,13 @@ define void @v_shuffle_v4p3_v2p3__3_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1161,14 +1148,13 @@ define void @v_shuffle_v4p3_v2p3__3_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -1309,16 +1295,15 @@ define void @v_shuffle_v4p3_v2p3__3_3_1_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:1]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1326,16 +1311,15 @@ define void @v_shuffle_v4p3_v2p3__3_3_1_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:1]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -1729,8 +1713,7 @@ define void @v_shuffle_v4p3_v2p3__3_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -1746,8 +1729,7 @@ define void @v_shuffle_v4p3_v2p3__3_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -1782,14 +1764,13 @@ define void @v_shuffle_v4p3_v2p3__3_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1799,14 +1780,13 @@ define void @v_shuffle_v4p3_v2p3__3_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -2158,8 +2138,7 @@ define void @v_shuffle_v4p3_v2p3__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -2172,8 +2151,7 @@ define void @v_shuffle_v4p3_v2p3__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -2252,16 +2230,15 @@ define void @v_shuffle_v4p3_v2p3__3_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:3]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2269,16 +2246,15 @@ define void @v_shuffle_v4p3_v2p3__3_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:3]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -2522,15 +2498,13 @@ define void @v_shuffle_v4p3_v2p3__3_3_1_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v5
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v1
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2540,15 +2514,13 @@ define void @v_shuffle_v4p3_v2p3__3_3_1_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -2832,16 +2804,15 @@ define void @v_shuffle_v4p3_v2p3__3_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:3]
+; GFX90A-NEXT: ; def v[0:1]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:5]
+; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2849,16 +2820,15 @@ define void @v_shuffle_v4p3_v2p3__3_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:3]
+; GFX940-NEXT: ; def v[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:5]
+; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -2944,8 +2914,7 @@ define void @v_shuffle_v4p3_v2p3__3_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -2958,8 +2927,7 @@ define void @v_shuffle_v4p3_v2p3__3_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -3645,9 +3613,9 @@ define void @s_shuffle_v4p3_v2p3__3_3_3_0() {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def s[4:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: s_mov_b32 s10, s9
; GFX90A-NEXT: s_mov_b32 s11, s4
+; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; use s[8:11]
; GFX90A-NEXT: ;;#ASMEND
@@ -3662,9 +3630,9 @@ define void @s_shuffle_v4p3_v2p3__3_3_3_0() {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def s[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: s_mov_b32 s10, s9
; GFX940-NEXT: s_mov_b32 s11, s0
+; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; use s[8:11]
; GFX940-NEXT: ;;#ASMEND
@@ -3715,35 +3683,19 @@ define void @s_shuffle_v4p3_v2p3__3_3_3_2() {
; GFX900-NEXT: ;;#ASMEND
; GFX900-NEXT: s_setpc_b64 s[30:31]
;
-; GFX90A-LABEL: s_shuffle_v4p3_v2p3__3_3_3_2:
-; GFX90A: ; %bb.0:
-; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def s[4:5]
-; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s5
-; GFX90A-NEXT: s_mov_b32 s9, s5
-; GFX90A-NEXT: s_mov_b32 s10, s5
-; GFX90A-NEXT: s_mov_b32 s11, s4
-; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; use s[8:11]
-; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_setpc_b64 s[30:31]
-;
-; GFX940-LABEL: s_shuffle_v4p3_v2p3__3_3_3_2:
-; GFX940: ; %bb.0:
-; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def s[0:1]
-; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s1
-; GFX940-NEXT: s_mov_b32 s9, s1
-; GFX940-NEXT: s_mov_b32 s10, s1
-; GFX940-NEXT: s_mov_b32 s11, s0
-; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; use s[8:11]
-; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_setpc_b64 s[30:31]
+; GFX90APLUS-LABEL: s_shuffle_v4p3_v2p3__3_3_3_2:
+; GFX90APLUS: ; %bb.0:
+; GFX90APLUS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90APLUS-NEXT: ;;#ASMSTART
+; GFX90APLUS-NEXT: ; def s[8:9]
+; GFX90APLUS-NEXT: ;;#ASMEND
+; GFX90APLUS-NEXT: s_mov_b32 s10, s9
+; GFX90APLUS-NEXT: s_mov_b32 s11, s8
+; GFX90APLUS-NEXT: s_mov_b32 s8, s9
+; GFX90APLUS-NEXT: ;;#ASMSTART
+; GFX90APLUS-NEXT: ; use s[8:11]
+; GFX90APLUS-NEXT: ;;#ASMEND
+; GFX90APLUS-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <2 x ptr addrspace(3)> asm "; def $0", "=s"()
%vec1 = call <2 x ptr addrspace(3)> asm "; def $0", "=s"()
%shuf = shufflevector <2 x ptr addrspace(3)> %vec0, <2 x ptr addrspace(3)> %vec1, <4 x i32> <i32 3, i32 3, i32 3, i32 2>
@@ -4186,9 +4138,9 @@ define void @s_shuffle_v4p3_v2p3__3_3_1_0() {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def s[4:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: s_mov_b32 s10, s5
; GFX90A-NEXT: s_mov_b32 s11, s4
+; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; use s[8:11]
; GFX90A-NEXT: ;;#ASMEND
@@ -4203,9 +4155,9 @@ define void @s_shuffle_v4p3_v2p3__3_3_1_0() {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def s[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: s_mov_b32 s10, s1
; GFX940-NEXT: s_mov_b32 s11, s0
+; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; use s[8:11]
; GFX940-NEXT: ;;#ASMEND
@@ -5089,15 +5041,14 @@ define void @s_shuffle_v4p3_v2p3__3_3_1_2() {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def s[4:5]
+; GFX90A-NEXT: ; def s[8:9]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def s[6:7]
+; GFX90A-NEXT: ; def s[4:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: s_mov_b32 s8, s7
-; GFX90A-NEXT: s_mov_b32 s9, s7
; GFX90A-NEXT: s_mov_b32 s10, s5
-; GFX90A-NEXT: s_mov_b32 s11, s6
+; GFX90A-NEXT: s_mov_b32 s11, s8
+; GFX90A-NEXT: s_mov_b32 s8, s9
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; use s[8:11]
; GFX90A-NEXT: ;;#ASMEND
@@ -5107,15 +5058,14 @@ define void @s_shuffle_v4p3_v2p3__3_3_1_2() {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def s[0:1]
+; GFX940-NEXT: ; def s[8:9]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def s[2:3]
+; GFX940-NEXT: ; def s[0:1]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_mov_b32 s8, s3
-; GFX940-NEXT: s_mov_b32 s9, s3
; GFX940-NEXT: s_mov_b32 s10, s1
-; GFX940-NEXT: s_mov_b32 s11, s2
+; GFX940-NEXT: s_mov_b32 s11, s8
+; GFX940-NEXT: s_mov_b32 s8, s9
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; use s[8:11]
; GFX940-NEXT: ;;#ASMEND
@@ -5490,5 +5440,3 @@ define void @s_shuffle_v4p3_v2p3__3_3_2_3() {
call void asm sideeffect "; use $0", "{s[8:11]}"(<4 x ptr addrspace(3)> %shuf)
ret void
}
-;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
-; GFX90APLUS: {{.*}}
diff --git a/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v3p3.ll b/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v3p3.ll
index d6bd8542831801..28bc61ce57815a 100644
--- a/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v3p3.ll
+++ b/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v3p3.ll
@@ -1280,8 +1280,7 @@ define void @v_shuffle_v4p3_v3p3__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
@@ -1296,8 +1295,7 @@ define void @v_shuffle_v4p3_v3p3__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
@@ -1435,8 +1433,7 @@ define void @v_shuffle_v4p3_v3p3__4_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
@@ -1453,8 +1450,8 @@ define void @v_shuffle_v4p3_v3p3__4_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
@@ -1964,15 +1961,14 @@ define void @v_shuffle_v4p3_v3p3__5_5_1_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v2
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1982,15 +1978,15 @@ define void @v_shuffle_v4p3_v3p3__5_5_1_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v0, v2
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -2142,17 +2138,16 @@ define void @v_shuffle_v4p3_v3p3__5_5_4_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v9, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[6:8]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v8
-; GFX90A-NEXT: v_mov_b32_e32 v1, v8
-; GFX90A-NEXT: v_mov_b32_e32 v2, v7
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v9, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2160,17 +2155,17 @@ define void @v_shuffle_v4p3_v3p3__5_5_4_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v9, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[6:8]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: v_mov_b32_e32 v0, v8
-; GFX940-NEXT: v_mov_b32_e32 v1, v8
-; GFX940-NEXT: v_mov_b32_e32 v2, v7
-; GFX940-NEXT: global_store_dwordx4 v9, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -3277,9 +3272,9 @@ define void @v_shuffle_v4p3_v3p3__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -3292,8 +3287,8 @@ define void @v_shuffle_v4p3_v3p3__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -3421,12 +3416,12 @@ define void @v_shuffle_v4p3_v3p3__4_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -3439,12 +3434,12 @@ define void @v_shuffle_v4p3_v3p3__4_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -3991,17 +3986,17 @@ define void @v_shuffle_v4p3_v3p3__5_5_1_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
-; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v5, 0
; GFX90A-NEXT: v_mov_b32_e32 v0, v2
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v7, 0
+; GFX90A-NEXT: ;;#ASMSTART
+; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v6
+; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4009,18 +4004,17 @@ define void @v_shuffle_v4p3_v3p3__5_5_1_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
-; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v5, 0
-; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_mov_b32_e32 v7, 0
; GFX940-NEXT: v_mov_b32_e32 v0, v2
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: ;;#ASMSTART
+; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ;;#ASMEND
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -4112,17 +4106,17 @@ define void @v_shuffle_v4p3_v3p3__5_5_4_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
+; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[6:8]
+; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v8
-; GFX90A-NEXT: v_mov_b32_e32 v1, v8
-; GFX90A-NEXT: v_mov_b32_e32 v2, v7
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v6
+; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4130,17 +4124,18 @@ define void @v_shuffle_v4p3_v3p3__5_5_4_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v5, 0
+; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[6:8]
+; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: v_mov_b32_e32 v0, v8
-; GFX940-NEXT: v_mov_b32_e32 v1, v8
-; GFX940-NEXT: v_mov_b32_e32 v2, v7
-; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v6
+; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -4317,8 +4312,7 @@ define void @v_shuffle_v4p3_v3p3__4_3_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:6]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
@@ -4333,8 +4327,7 @@ define void @v_shuffle_v4p3_v3p3__4_3_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:6]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
@@ -4858,17 +4851,16 @@ define void @v_shuffle_v4p3_v3p3__5_5_1_3(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v6
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4876,17 +4868,17 @@ define void @v_shuffle_v4p3_v3p3__5_5_1_3(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v6
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -4972,31 +4964,28 @@ define void @v_shuffle_v4p3_v3p3__5_5_4_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4p3_v3p3__5_5_4_3:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v6
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4p3_v3p3__5_5_4_3:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v6
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -6094,9 +6083,9 @@ define void @v_shuffle_v4p3_v3p3__1_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v5, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6113,8 +6102,8 @@ define void @v_shuffle_v4p3_v3p3__1_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -6252,9 +6241,9 @@ define void @v_shuffle_v4p3_v3p3__4_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v0, v2
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6267,8 +6256,8 @@ define void @v_shuffle_v4p3_v3p3__4_5_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_mov_b32_e32 v0, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -6727,17 +6716,16 @@ define void @v_shuffle_v4p3_v3p3__5_5_1_5(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[2:4]
+; GFX90A-NEXT: ; def v[0:2]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v7, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:6]
+; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v6
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v6
-; GFX90A-NEXT: global_store_dwordx4 v7, v[0:3], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -6745,17 +6733,17 @@ define void @v_shuffle_v4p3_v3p3__5_5_1_5(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[2:4]
+; GFX940-NEXT: ; def v[0:2]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, 0
+; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:6]
+; GFX940-NEXT: ; def v[2:4]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v6
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
-; GFX940-NEXT: v_mov_b32_e32 v3, v6
-; GFX940-NEXT: global_store_dwordx4 v7, v[0:3], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <3 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -6893,11 +6881,10 @@ define void @v_shuffle_v4p3_v3p3__5_5_4_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:4]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v5, 0
; GFX90A-NEXT: v_mov_b32_e32 v0, v4
+; GFX90A-NEXT: v_mov_b32_e32 v5, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v1, v4
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v5, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -6910,9 +6897,8 @@ define void @v_shuffle_v4p3_v3p3__5_5_4_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v5, 0
; GFX940-NEXT: v_mov_b32_e32 v0, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v1, v4
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v5, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
diff --git a/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v4p3.ll b/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v4p3.ll
index e4c7cebb231bd8..9cc1b9fe6cf0e0 100644
--- a/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v4p3.ll
+++ b/llvm/test/CodeGen/AMDGPU/shufflevector.v4p3.v4p3.ll
@@ -343,13 +343,12 @@ define void @v_shuffle_v4p3_v4p3__7_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
-; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -364,9 +363,8 @@ define void @v_shuffle_v4p3_v4p3__7_0_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
-; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -456,8 +454,7 @@ define void @v_shuffle_v4p3_v4p3__7_2_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -472,8 +469,8 @@ define void @v_shuffle_v4p3_v4p3__7_2_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -560,9 +557,8 @@ define void @v_shuffle_v4p3_v4p3__7_4_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
-; GFX90A-NEXT: global_store_dwordx4 v4, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -573,9 +569,8 @@ define void @v_shuffle_v4p3_v4p3__7_4_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
-; GFX940-NEXT: global_store_dwordx4 v4, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -649,8 +644,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_u_u(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -662,8 +656,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_u_u(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -1169,11 +1162,10 @@ define void @v_shuffle_v4p3_v4p3__7_7_7_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v5
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1188,11 +1180,10 @@ define void @v_shuffle_v4p3_v4p3__7_7_7_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v5
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -1287,17 +1278,16 @@ define void @v_shuffle_v4p3_v4p3__7_7_7_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v7
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -1305,17 +1295,17 @@ define void @v_shuffle_v4p3_v4p3__7_7_7_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v7
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -1402,31 +1392,26 @@ define void @v_shuffle_v4p3_v4p3__7_7_7_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4p3_v4p3__7_7_7_4:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4p3_v4p3__7_7_7_4:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -1508,31 +1493,26 @@ define void @v_shuffle_v4p3_v4p3__7_7_7_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4p3_v4p3__7_7_7_6:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4p3_v4p3__7_7_7_6:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -1709,8 +1689,7 @@ define void @v_shuffle_v4p3_v4p3__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -1724,8 +1703,7 @@ define void @v_shuffle_v4p3_v4p3__1_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -1809,8 +1787,7 @@ define void @v_shuffle_v4p3_v4p3__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -1824,8 +1801,7 @@ define void @v_shuffle_v4p3_v4p3__3_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -1914,8 +1890,7 @@ define void @v_shuffle_v4p3_v4p3__5_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -1933,8 +1908,7 @@ define void @v_shuffle_v4p3_v4p3__5_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -2038,8 +2012,7 @@ define void @v_shuffle_v4p3_v4p3__7_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -2057,8 +2030,7 @@ define void @v_shuffle_v4p3_v4p3__7_0_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -2214,17 +2186,16 @@ define void @v_shuffle_v4p3_v4p3__7_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v0
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v8, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v0
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: global_store_dwordx4 v8, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2232,17 +2203,16 @@ define void @v_shuffle_v4p3_v4p3__7_2_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v0
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: global_store_dwordx4 v8, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -2339,12 +2309,11 @@ define void @v_shuffle_v4p3_v4p3__7_4_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v0
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v0
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2354,15 +2323,15 @@ define void @v_shuffle_v4p3_v4p3__7_4_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -2460,8 +2429,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_0_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -2479,8 +2447,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_0_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -2643,11 +2610,10 @@ define void @v_shuffle_v4p3_v4p3__7_7_1_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v5
-; GFX90A-NEXT: v_mov_b32_e32 v4, v1
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2662,11 +2628,10 @@ define void @v_shuffle_v4p3_v4p3__7_7_1_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v5
-; GFX940-NEXT: v_mov_b32_e32 v4, v1
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -2759,17 +2724,16 @@ define void @v_shuffle_v4p3_v4p3__7_7_3_0(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2777,17 +2741,16 @@ define void @v_shuffle_v4p3_v4p3__7_7_3_0(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -2883,11 +2846,11 @@ define void @v_shuffle_v4p3_v4p3__7_7_5_0(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -2897,14 +2860,15 @@ define void @v_shuffle_v4p3_v4p3__7_7_5_0(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -3517,8 +3481,7 @@ define void @v_shuffle_v4p3_v4p3__7_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v1
; GFX90A-NEXT: v_mov_b32_e32 v5, v1
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -3536,8 +3499,7 @@ define void @v_shuffle_v4p3_v4p3__7_0_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v1
; GFX940-NEXT: v_mov_b32_e32 v5, v1
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -3573,17 +3535,16 @@ define void @v_shuffle_v4p3_v4p3__7_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v1
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v8, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v1
+; GFX90A-NEXT: v_mov_b32_e32 v5, v1
+; GFX90A-NEXT: global_store_dwordx4 v8, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -3591,17 +3552,16 @@ define void @v_shuffle_v4p3_v4p3__7_2_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v1
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[6:7], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v1
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: global_store_dwordx4 v8, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -3697,12 +3657,11 @@ define void @v_shuffle_v4p3_v4p3__7_4_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v1
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v1
+; GFX90A-NEXT: v_mov_b32_e32 v5, v1
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -3712,15 +3671,15 @@ define void @v_shuffle_v4p3_v4p3__7_4_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v1
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v1
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -3818,8 +3777,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_1_1(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v1
; GFX90A-NEXT: v_mov_b32_e32 v5, v1
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -3837,8 +3795,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_1_1(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v1
; GFX940-NEXT: v_mov_b32_e32 v5, v1
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -4444,8 +4401,7 @@ define void @v_shuffle_v4p3_v4p3__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4458,8 +4414,7 @@ define void @v_shuffle_v4p3_v4p3__1_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4540,8 +4495,7 @@ define void @v_shuffle_v4p3_v4p3__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4554,8 +4508,7 @@ define void @v_shuffle_v4p3_v4p3__3_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4639,8 +4592,7 @@ define void @v_shuffle_v4p3_v4p3__5_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v5
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4656,8 +4608,8 @@ define void @v_shuffle_v4p3_v4p3__5_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4755,8 +4707,7 @@ define void @v_shuffle_v4p3_v4p3__7_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -4772,8 +4723,8 @@ define void @v_shuffle_v4p3_v4p3__7_2_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -4863,17 +4814,15 @@ define void @v_shuffle_v4p3_v4p3__7_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v2
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -4881,17 +4830,15 @@ define void @v_shuffle_v4p3_v4p3__7_0_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -5042,8 +4989,7 @@ define void @v_shuffle_v4p3_v4p3__7_4_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -5060,8 +5006,7 @@ define void @v_shuffle_v4p3_v4p3__7_4_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -5158,8 +5103,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_2_2(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -5176,8 +5120,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_2_2(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v3, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -5388,17 +5331,16 @@ define void @v_shuffle_v4p3_v4p3__7_7_1_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -5406,17 +5348,16 @@ define void @v_shuffle_v4p3_v4p3__7_7_1_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -5449,17 +5390,16 @@ define void @v_shuffle_v4p3_v4p3__7_7_3_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v7
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -5467,17 +5407,16 @@ define void @v_shuffle_v4p3_v4p3__7_7_3_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v7
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -5568,16 +5507,16 @@ define void @v_shuffle_v4p3_v4p3__7_7_5_2(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v10, 0
+; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, v7
-; GFX90A-NEXT: v_mov_b32_e32 v8, v5
-; GFX90A-NEXT: v_mov_b32_e32 v9, v2
-; GFX90A-NEXT: global_store_dwordx4 v10, v[6:9], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v7
+; GFX90A-NEXT: v_mov_b32_e32 v1, v7
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -5585,16 +5524,17 @@ define void @v_shuffle_v4p3_v4p3__7_7_5_2(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v10, 0
+; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v7
-; GFX940-NEXT: v_mov_b32_e32 v8, v5
-; GFX940-NEXT: v_mov_b32_e32 v9, v2
-; GFX940-NEXT: global_store_dwordx4 v10, v[6:9], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: v_mov_b32_e32 v1, v7
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -6193,17 +6133,15 @@ define void @v_shuffle_v4p3_v4p3__7_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[4:7]
+; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
-; GFX90A-NEXT: ; def v[0:3]
+; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -6211,17 +6149,15 @@ define void @v_shuffle_v4p3_v4p3__7_0_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[4:7]
+; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
-; GFX940-NEXT: ; def v[0:3]
+; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -6314,8 +6250,7 @@ define void @v_shuffle_v4p3_v4p3__7_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6331,8 +6266,8 @@ define void @v_shuffle_v4p3_v4p3__7_2_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -6372,8 +6307,7 @@ define void @v_shuffle_v4p3_v4p3__7_4_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6390,8 +6324,7 @@ define void @v_shuffle_v4p3_v4p3__7_4_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v4
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[4:5] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -6488,8 +6421,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_3_3(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v6
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -6506,8 +6438,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_3_3(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
-; GFX940-NEXT: v_mov_b32_e32 v1, v6
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[6:7] op_sel:[1,0]
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -7195,8 +7126,7 @@ define void @v_shuffle_v4p3_v4p3__5_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -7210,8 +7140,7 @@ define void @v_shuffle_v4p3_v4p3__5_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -7297,8 +7226,7 @@ define void @v_shuffle_v4p3_v4p3__7_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v0
; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -7312,8 +7240,7 @@ define void @v_shuffle_v4p3_v4p3__7_4_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v0
; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -7404,8 +7331,7 @@ define void @v_shuffle_v4p3_v4p3__7_0_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v6, v2
; GFX90A-NEXT: v_mov_b32_e32 v7, v2
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
@@ -7423,8 +7349,7 @@ define void @v_shuffle_v4p3_v4p3__7_0_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v6, v2
; GFX940-NEXT: v_mov_b32_e32 v7, v2
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
@@ -7523,8 +7448,7 @@ define void @v_shuffle_v4p3_v4p3__7_2_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v4
; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
@@ -7541,8 +7465,8 @@ define void @v_shuffle_v4p3_v4p3__7_2_4_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v4
; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
@@ -7686,31 +7610,28 @@ define void @v_shuffle_v4p3_v4p3__7_6_4_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4p3_v4p3__7_6_4_4:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v0
-; GFX90A-NEXT: v_mov_b32_e32 v7, v0
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v0
+; GFX90A-NEXT: v_mov_b32_e32 v5, v0
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4p3_v4p3__7_6_4_4:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v0
-; GFX940-NEXT: v_mov_b32_e32 v7, v0
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v0
+; GFX940-NEXT: v_mov_b32_e32 v5, v0
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -7904,11 +7825,11 @@ define void @v_shuffle_v4p3_v4p3__7_7_1_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -7918,14 +7839,15 @@ define void @v_shuffle_v4p3_v4p3__7_7_1_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -8022,10 +7944,9 @@ define void @v_shuffle_v4p3_v4p3__7_7_3_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v7
; GFX90A-NEXT: v_mov_b32_e32 v1, v7
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -8040,10 +7961,10 @@ define void @v_shuffle_v4p3_v4p3__7_7_3_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v7
; GFX940-NEXT: v_mov_b32_e32 v1, v7
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -8076,9 +7997,8 @@ define void @v_shuffle_v4p3_v4p3__7_7_5_4(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v4, v1
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -8090,9 +8010,8 @@ define void @v_shuffle_v4p3_v4p3__7_7_5_4(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v4, v1
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -8703,12 +8622,10 @@ define void @v_shuffle_v4p3_v4p3__7_0_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
-; GFX90A-NEXT: v_mov_b32_e32 v4, v5
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -8718,16 +8635,14 @@ define void @v_shuffle_v4p3_v4p3__7_0_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v5
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -8822,8 +8737,7 @@ define void @v_shuffle_v4p3_v4p3__7_2_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v0, v7
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v5
; GFX90A-NEXT: v_mov_b32_e32 v3, v5
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
@@ -8840,8 +8754,8 @@ define void @v_shuffle_v4p3_v4p3__7_2_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
-; GFX940-NEXT: v_mov_b32_e32 v0, v7
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v5
; GFX940-NEXT: v_mov_b32_e32 v3, v5
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
@@ -8938,8 +8852,7 @@ define void @v_shuffle_v4p3_v4p3__7_4_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v1
; GFX90A-NEXT: v_mov_b32_e32 v5, v1
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
@@ -8953,8 +8866,7 @@ define void @v_shuffle_v4p3_v4p3__7_4_5_5(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v6, 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v1
; GFX940-NEXT: v_mov_b32_e32 v5, v1
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
@@ -8985,31 +8897,28 @@ define void @v_shuffle_v4p3_v4p3__7_6_5_5(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4p3_v4p3__7_6_5_5:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v1
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v4, v1
+; GFX90A-NEXT: v_mov_b32_e32 v5, v1
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4p3_v4p3__7_6_5_5:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v1
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v4, v1
+; GFX940-NEXT: v_mov_b32_e32 v5, v1
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -9584,8 +9493,7 @@ define void @v_shuffle_v4p3_v4p3__1_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v1
-; GFX90A-NEXT: v_mov_b32_e32 v3, v4
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v5, v4
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9602,8 +9510,7 @@ define void @v_shuffle_v4p3_v4p3__1_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v1
-; GFX940-NEXT: v_mov_b32_e32 v3, v4
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v5, v4
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9701,8 +9608,7 @@ define void @v_shuffle_v4p3_v4p3__3_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v6
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v7, v6
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9718,8 +9624,8 @@ define void @v_shuffle_v4p3_v4p3__3_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v5, v6
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[6:7] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v7, v6
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9799,8 +9705,7 @@ define void @v_shuffle_v4p3_v4p3__5_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v1
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9813,8 +9718,7 @@ define void @v_shuffle_v4p3_v4p3__5_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v1
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9897,8 +9801,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v3, v2
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -9911,8 +9814,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v3, v2
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -9998,8 +9900,7 @@ define void @v_shuffle_v4p3_v4p3__7_0_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v5, v4
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -10016,8 +9917,7 @@ define void @v_shuffle_v4p3_v4p3__7_0_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v5, v4
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -10116,8 +10016,7 @@ define void @v_shuffle_v4p3_v4p3__7_2_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v7, v6
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -10133,8 +10032,8 @@ define void @v_shuffle_v4p3_v4p3__7_2_6_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v7, v6
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -10222,31 +10121,26 @@ define void @v_shuffle_v4p3_v4p3__7_4_6_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4p3_v4p3__7_4_6_6:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v2
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v3, v2
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4p3_v4p3__7_4_6_6:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v2
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v3, v2
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -10486,15 +10380,14 @@ define void @v_shuffle_v4p3_v4p3__7_7_1_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v10, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v6, v5
-; GFX90A-NEXT: v_mov_b32_e32 v7, v5
-; GFX90A-NEXT: v_mov_b32_e32 v8, v1
-; GFX90A-NEXT: v_mov_b32_e32 v9, v4
-; GFX90A-NEXT: global_store_dwordx4 v10, v[6:9], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v0, v5
+; GFX90A-NEXT: v_mov_b32_e32 v1, v5
+; GFX90A-NEXT: global_store_dwordx4 v6, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
@@ -10504,15 +10397,15 @@ define void @v_shuffle_v4p3_v4p3__7_7_1_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v10, 0
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v8, v1
-; GFX940-NEXT: v_mov_b32_e32 v6, v5
-; GFX940-NEXT: v_mov_b32_e32 v7, v5
-; GFX940-NEXT: v_mov_b32_e32 v9, v4
-; GFX940-NEXT: global_store_dwordx4 v10, v[6:9], s[0:1] sc0 sc1
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[0:1], v[4:5] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v0, v5
+; GFX940-NEXT: v_mov_b32_e32 v1, v5
+; GFX940-NEXT: global_store_dwordx4 v6, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -10610,10 +10503,9 @@ define void @v_shuffle_v4p3_v4p3__7_7_3_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[4:7]
; GFX90A-NEXT: ;;#ASMEND
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[6:7] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v0, v7
; GFX90A-NEXT: v_mov_b32_e32 v1, v7
-; GFX90A-NEXT: v_mov_b32_e32 v2, v3
-; GFX90A-NEXT: v_mov_b32_e32 v3, v6
; GFX90A-NEXT: global_store_dwordx4 v8, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -10628,10 +10520,10 @@ define void @v_shuffle_v4p3_v4p3__7_7_3_6(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[4:7]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[2:3], v[6:7] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v0, v7
; GFX940-NEXT: v_mov_b32_e32 v1, v7
-; GFX940-NEXT: v_mov_b32_e32 v3, v6
; GFX940-NEXT: global_store_dwordx4 v8, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -10714,31 +10606,26 @@ define void @v_shuffle_v4p3_v4p3__7_7_5_6(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4p3_v4p3__7_7_5_6:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v3
-; GFX90A-NEXT: v_mov_b32_e32 v6, v1
-; GFX90A-NEXT: v_mov_b32_e32 v7, v2
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v6, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[2:3] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4p3_v4p3__7_7_5_6:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v3
-; GFX940-NEXT: v_mov_b32_e32 v6, v1
-; GFX940-NEXT: v_mov_b32_e32 v7, v2
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v6, 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[2:3] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -11246,8 +11133,7 @@ define void @v_shuffle_v4p3_v4p3__7_0_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[2:5]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v6, 0
-; GFX90A-NEXT: v_mov_b32_e32 v2, v5
-; GFX90A-NEXT: v_mov_b32_e32 v3, v0
+; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v4, v5
; GFX90A-NEXT: global_store_dwordx4 v6, v[2:5], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -11264,8 +11150,7 @@ define void @v_shuffle_v4p3_v4p3__7_0_7_7(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[2:5]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v2, v5
-; GFX940-NEXT: v_mov_b32_e32 v3, v0
+; GFX940-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[0:1] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v4, v5
; GFX940-NEXT: global_store_dwordx4 v6, v[2:5], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -11364,8 +11249,7 @@ define void @v_shuffle_v4p3_v4p3__7_2_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v7
-; GFX90A-NEXT: v_mov_b32_e32 v5, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v6, v7
; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -11381,8 +11265,8 @@ define void @v_shuffle_v4p3_v4p3__7_2_7_7(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: v_mov_b32_e32 v4, v7
-; GFX940-NEXT: v_mov_b32_e32 v5, v2
+; GFX940-NEXT: s_nop 0
+; GFX940-NEXT: v_pk_mov_b32 v[4:5], v[6:7], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v6, v7
; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
@@ -11471,31 +11355,26 @@ define void @v_shuffle_v4p3_v4p3__7_4_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-LABEL: v_shuffle_v4p3_v4p3__7_4_7_7:
; GFX90A: ; %bb.0:
; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX90A-NEXT: v_mov_b32_e32 v8, 0
; GFX90A-NEXT: ;;#ASMSTART
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
-; GFX90A-NEXT: v_mov_b32_e32 v4, v3
-; GFX90A-NEXT: v_mov_b32_e32 v5, v0
-; GFX90A-NEXT: v_mov_b32_e32 v6, v3
-; GFX90A-NEXT: v_mov_b32_e32 v7, v3
-; GFX90A-NEXT: global_store_dwordx4 v8, v[4:7], s[16:17]
+; GFX90A-NEXT: v_mov_b32_e32 v4, 0
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX90A-NEXT: v_mov_b32_e32 v2, v3
+; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
; GFX90A-NEXT: s_setpc_b64 s[30:31]
;
; GFX940-LABEL: v_shuffle_v4p3_v4p3__7_4_7_7:
; GFX940: ; %bb.0:
; GFX940-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT: v_mov_b32_e32 v8, 0
; GFX940-NEXT: ;;#ASMSTART
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
-; GFX940-NEXT: s_nop 0
-; GFX940-NEXT: v_mov_b32_e32 v4, v3
-; GFX940-NEXT: v_mov_b32_e32 v5, v0
-; GFX940-NEXT: v_mov_b32_e32 v6, v3
-; GFX940-NEXT: v_mov_b32_e32 v7, v3
-; GFX940-NEXT: global_store_dwordx4 v8, v[4:7], s[0:1] sc0 sc1
+; GFX940-NEXT: v_mov_b32_e32 v4, 0
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
+; GFX940-NEXT: v_mov_b32_e32 v2, v3
+; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
; GFX940-NEXT: s_setpc_b64 s[30:31]
%vec0 = call <4 x ptr addrspace(3)> asm "; def $0", "=v"()
@@ -11573,8 +11452,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_7_7(ptr addrspace(1) inreg %ptr) {
; GFX90A-NEXT: ; def v[0:3]
; GFX90A-NEXT: ;;#ASMEND
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
-; GFX90A-NEXT: v_mov_b32_e32 v0, v3
-; GFX90A-NEXT: v_mov_b32_e32 v1, v2
+; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX90A-NEXT: v_mov_b32_e32 v2, v3
; GFX90A-NEXT: global_store_dwordx4 v4, v[0:3], s[16:17]
; GFX90A-NEXT: s_waitcnt vmcnt(0)
@@ -11587,8 +11465,7 @@ define void @v_shuffle_v4p3_v4p3__7_6_7_7(ptr addrspace(1) inreg %ptr) {
; GFX940-NEXT: ; def v[0:3]
; GFX940-NEXT: ;;#ASMEND
; GFX940-NEXT: v_mov_b32_e32 v4, 0
-; GFX940-NEXT: v_mov_b32_e32 v0, v3
-; GFX940-NEXT: v_mov_b32_e32 v1, v2
+; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
; GFX940-NEXT: v_mov_b32_e32 v2, v3
; GFX940-NEXT: global_store_dwordx4 v4, v[0:3], s[0:1] sc0 sc1
; GFX940-NEXT: s_waitcnt vmcnt(0)
More information about the llvm-commits
mailing list