Skip to content

Commit 84c8a20

Browse files
arsenmarsenm
committed
Enable op_sel_hi to avoid printing it
1 parent a6598e0 commit 84c8a20

13 files changed

+317
-312
lines changed

‎llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp‎

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -547,6 +547,11 @@ void AMDGPUDAGToDAGISel::SelectVectorShuffle(SDNode *N) {
547547
unsigned Src1OpSel =
548548
Src1SubReg == AMDGPU::sub1 ? SISrcMods::OP_SEL_0 : SISrcMods::NONE;
549549

550+
// Enable op_sel_hi to avoid printing it. This should have no effect on the
551+
// result.
552+
Src0OpSel |= SISrcMods::OP_SEL_1;
553+
Src1OpSel |= SISrcMods::OP_SEL_1;
554+
550555
SDValue Src0OpSelVal = CurDAG->getTargetConstant(Src0OpSel, DL, MVT::i32);
551556
SDValue Src1OpSelVal = CurDAG->getTargetConstant(Src1OpSel, DL, MVT::i32);
552557
SDValue ZeroMods = CurDAG->getTargetConstant(0, DL, MVT::i32);

‎llvm/test/CodeGen/AMDGPU/shufflevector.v2f32.v2f32.ll‎

Lines changed: 6 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -177,7 +177,7 @@ define void @v_shuffle_v2f32_v2f32__3_0(ptr addrspace(1) inreg %ptr) {
177177
; GFX90A-NEXT: ;;#ASMSTART
178178
; GFX90A-NEXT: ; def v[2:3]
179179
; GFX90A-NEXT: ;;#ASMEND
180-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
180+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
181181
; GFX90A-NEXT: global_store_dwordx2 v4, v[0:1], s[16:17]
182182
; GFX90A-NEXT: s_waitcnt vmcnt(0)
183183
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -193,7 +193,7 @@ define void @v_shuffle_v2f32_v2f32__3_0(ptr addrspace(1) inreg %ptr) {
193193
; GFX940-NEXT: ; def v[2:3]
194194
; GFX940-NEXT: ;;#ASMEND
195195
; GFX940-NEXT: s_nop 0
196-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
196+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
197197
; GFX940-NEXT: global_store_dwordx2 v4, v[0:1], s[0:1] sc0 sc1
198198
; GFX940-NEXT: s_waitcnt vmcnt(0)
199199
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -277,7 +277,7 @@ define void @v_shuffle_v2f32_v2f32__3_2(ptr addrspace(1) inreg %ptr) {
277277
; GFX90A-NEXT: ; def v[0:1]
278278
; GFX90A-NEXT: ;;#ASMEND
279279
; GFX90A-NEXT: v_mov_b32_e32 v2, 0
280-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
280+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
281281
; GFX90A-NEXT: global_store_dwordx2 v2, v[0:1], s[16:17]
282282
; GFX90A-NEXT: s_waitcnt vmcnt(0)
283283
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -289,7 +289,7 @@ define void @v_shuffle_v2f32_v2f32__3_2(ptr addrspace(1) inreg %ptr) {
289289
; GFX940-NEXT: ; def v[0:1]
290290
; GFX940-NEXT: ;;#ASMEND
291291
; GFX940-NEXT: v_mov_b32_e32 v2, 0
292-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
292+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
293293
; GFX940-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1] sc0 sc1
294294
; GFX940-NEXT: s_waitcnt vmcnt(0)
295295
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -447,7 +447,7 @@ define void @v_shuffle_v2f32_v2f32__1_0(ptr addrspace(1) inreg %ptr) {
447447
; GFX90A-NEXT: ; def v[0:1]
448448
; GFX90A-NEXT: ;;#ASMEND
449449
; GFX90A-NEXT: v_mov_b32_e32 v2, 0
450-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
450+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
451451
; GFX90A-NEXT: global_store_dwordx2 v2, v[0:1], s[16:17]
452452
; GFX90A-NEXT: s_waitcnt vmcnt(0)
453453
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -459,7 +459,7 @@ define void @v_shuffle_v2f32_v2f32__1_0(ptr addrspace(1) inreg %ptr) {
459459
; GFX940-NEXT: ; def v[0:1]
460460
; GFX940-NEXT: ;;#ASMEND
461461
; GFX940-NEXT: v_mov_b32_e32 v2, 0
462-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
462+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
463463
; GFX940-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1] sc0 sc1
464464
; GFX940-NEXT: s_waitcnt vmcnt(0)
465465
; GFX940-NEXT: s_setpc_b64 s[30:31]

‎llvm/test/CodeGen/AMDGPU/shufflevector.v2f32.v3f32.ll‎

Lines changed: 6 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -633,7 +633,7 @@ define void @v_shuffle_v2f32_v3f32__1_0(ptr addrspace(1) inreg %ptr) {
633633
; GFX90A-NEXT: ; def v[0:2]
634634
; GFX90A-NEXT: ;;#ASMEND
635635
; GFX90A-NEXT: v_mov_b32_e32 v3, 0
636-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
636+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
637637
; GFX90A-NEXT: global_store_dwordx2 v3, v[0:1], s[16:17]
638638
; GFX90A-NEXT: s_waitcnt vmcnt(0)
639639
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -645,7 +645,7 @@ define void @v_shuffle_v2f32_v3f32__1_0(ptr addrspace(1) inreg %ptr) {
645645
; GFX940-NEXT: ; def v[0:2]
646646
; GFX940-NEXT: ;;#ASMEND
647647
; GFX940-NEXT: v_mov_b32_e32 v3, 0
648-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
648+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
649649
; GFX940-NEXT: global_store_dwordx2 v3, v[0:1], s[0:1] sc0 sc1
650650
; GFX940-NEXT: s_waitcnt vmcnt(0)
651651
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -767,7 +767,7 @@ define void @v_shuffle_v2f32_v3f32__4_0(ptr addrspace(1) inreg %ptr) {
767767
; GFX90A-NEXT: ;;#ASMSTART
768768
; GFX90A-NEXT: ; def v[2:4]
769769
; GFX90A-NEXT: ;;#ASMEND
770-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
770+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
771771
; GFX90A-NEXT: global_store_dwordx2 v5, v[0:1], s[16:17]
772772
; GFX90A-NEXT: s_waitcnt vmcnt(0)
773773
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -783,7 +783,7 @@ define void @v_shuffle_v2f32_v3f32__4_0(ptr addrspace(1) inreg %ptr) {
783783
; GFX940-NEXT: ; def v[2:4]
784784
; GFX940-NEXT: ;;#ASMEND
785785
; GFX940-NEXT: s_nop 0
786-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
786+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
787787
; GFX940-NEXT: global_store_dwordx2 v5, v[0:1], s[0:1] sc0 sc1
788788
; GFX940-NEXT: s_waitcnt vmcnt(0)
789789
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -1477,7 +1477,7 @@ define void @v_shuffle_v2f32_v3f32__4_3(ptr addrspace(1) inreg %ptr) {
14771477
; GFX90A-NEXT: ; def v[0:2]
14781478
; GFX90A-NEXT: ;;#ASMEND
14791479
; GFX90A-NEXT: v_mov_b32_e32 v3, 0
1480-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
1480+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
14811481
; GFX90A-NEXT: global_store_dwordx2 v3, v[0:1], s[16:17]
14821482
; GFX90A-NEXT: s_waitcnt vmcnt(0)
14831483
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -1489,7 +1489,7 @@ define void @v_shuffle_v2f32_v3f32__4_3(ptr addrspace(1) inreg %ptr) {
14891489
; GFX940-NEXT: ; def v[0:2]
14901490
; GFX940-NEXT: ;;#ASMEND
14911491
; GFX940-NEXT: v_mov_b32_e32 v3, 0
1492-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
1492+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
14931493
; GFX940-NEXT: global_store_dwordx2 v3, v[0:1], s[0:1] sc0 sc1
14941494
; GFX940-NEXT: s_waitcnt vmcnt(0)
14951495
; GFX940-NEXT: s_setpc_b64 s[30:31]

‎llvm/test/CodeGen/AMDGPU/shufflevector.v2f32.v4f32.ll‎

Lines changed: 20 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -339,7 +339,7 @@ define void @v_shuffle_v2f32_v4f32__7_0(ptr addrspace(1) inreg %ptr) {
339339
; GFX90A-NEXT: ;;#ASMSTART
340340
; GFX90A-NEXT: ; def v[2:5]
341341
; GFX90A-NEXT: ;;#ASMEND
342-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
342+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
343343
; GFX90A-NEXT: global_store_dwordx2 v6, v[0:1], s[16:17]
344344
; GFX90A-NEXT: s_waitcnt vmcnt(0)
345345
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -355,7 +355,7 @@ define void @v_shuffle_v2f32_v4f32__7_0(ptr addrspace(1) inreg %ptr) {
355355
; GFX940-NEXT: ; def v[2:5]
356356
; GFX940-NEXT: ;;#ASMEND
357357
; GFX940-NEXT: s_nop 0
358-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
358+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[0:1] op_sel:[1,0]
359359
; GFX940-NEXT: global_store_dwordx2 v6, v[0:1], s[0:1] sc0 sc1
360360
; GFX940-NEXT: s_waitcnt vmcnt(0)
361361
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -445,7 +445,7 @@ define void @v_shuffle_v2f32_v4f32__7_2(ptr addrspace(1) inreg %ptr) {
445445
; GFX90A-NEXT: ;;#ASMSTART
446446
; GFX90A-NEXT: ; def v[4:7]
447447
; GFX90A-NEXT: ;;#ASMEND
448-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0] op_sel_hi:[0,0]
448+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
449449
; GFX90A-NEXT: global_store_dwordx2 v8, v[0:1], s[16:17]
450450
; GFX90A-NEXT: s_waitcnt vmcnt(0)
451451
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -461,7 +461,7 @@ define void @v_shuffle_v2f32_v4f32__7_2(ptr addrspace(1) inreg %ptr) {
461461
; GFX940-NEXT: ; def v[4:7]
462462
; GFX940-NEXT: ;;#ASMEND
463463
; GFX940-NEXT: s_nop 0
464-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0] op_sel_hi:[0,0]
464+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[6:7], v[2:3] op_sel:[1,0]
465465
; GFX940-NEXT: global_store_dwordx2 v8, v[0:1], s[0:1] sc0 sc1
466466
; GFX940-NEXT: s_waitcnt vmcnt(0)
467467
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -634,7 +634,7 @@ define void @v_shuffle_v2f32_v4f32__7_6(ptr addrspace(1) inreg %ptr) {
634634
; GFX90A-NEXT: ; def v[0:3]
635635
; GFX90A-NEXT: ;;#ASMEND
636636
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
637-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0] op_sel_hi:[0,0]
637+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
638638
; GFX90A-NEXT: global_store_dwordx2 v4, v[0:1], s[16:17]
639639
; GFX90A-NEXT: s_waitcnt vmcnt(0)
640640
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -646,7 +646,7 @@ define void @v_shuffle_v2f32_v4f32__7_6(ptr addrspace(1) inreg %ptr) {
646646
; GFX940-NEXT: ; def v[0:3]
647647
; GFX940-NEXT: ;;#ASMEND
648648
; GFX940-NEXT: v_mov_b32_e32 v4, 0
649-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0] op_sel_hi:[0,0]
649+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
650650
; GFX940-NEXT: global_store_dwordx2 v4, v[0:1], s[0:1] sc0 sc1
651651
; GFX940-NEXT: s_waitcnt vmcnt(0)
652652
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -804,7 +804,7 @@ define void @v_shuffle_v2f32_v4f32__1_0(ptr addrspace(1) inreg %ptr) {
804804
; GFX90A-NEXT: ; def v[0:3]
805805
; GFX90A-NEXT: ;;#ASMEND
806806
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
807-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
807+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
808808
; GFX90A-NEXT: global_store_dwordx2 v4, v[0:1], s[16:17]
809809
; GFX90A-NEXT: s_waitcnt vmcnt(0)
810810
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -816,7 +816,7 @@ define void @v_shuffle_v2f32_v4f32__1_0(ptr addrspace(1) inreg %ptr) {
816816
; GFX940-NEXT: ; def v[0:3]
817817
; GFX940-NEXT: ;;#ASMEND
818818
; GFX940-NEXT: v_mov_b32_e32 v4, 0
819-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
819+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
820820
; GFX940-NEXT: global_store_dwordx2 v4, v[0:1], s[0:1] sc0 sc1
821821
; GFX940-NEXT: s_waitcnt vmcnt(0)
822822
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -981,7 +981,7 @@ define void @v_shuffle_v2f32_v4f32__5_0(ptr addrspace(1) inreg %ptr) {
981981
; GFX90A-NEXT: ;;#ASMSTART
982982
; GFX90A-NEXT: ; def v[2:5]
983983
; GFX90A-NEXT: ;;#ASMEND
984-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
984+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
985985
; GFX90A-NEXT: global_store_dwordx2 v6, v[0:1], s[16:17]
986986
; GFX90A-NEXT: s_waitcnt vmcnt(0)
987987
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -997,7 +997,7 @@ define void @v_shuffle_v2f32_v4f32__5_0(ptr addrspace(1) inreg %ptr) {
997997
; GFX940-NEXT: ; def v[2:5]
998998
; GFX940-NEXT: ;;#ASMEND
999999
; GFX940-NEXT: s_nop 0
1000-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
1000+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[0:1] op_sel:[1,0]
10011001
; GFX940-NEXT: global_store_dwordx2 v6, v[0:1], s[0:1] sc0 sc1
10021002
; GFX940-NEXT: s_waitcnt vmcnt(0)
10031003
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -1598,7 +1598,7 @@ define void @v_shuffle_v2f32_v4f32__3_2(ptr addrspace(1) inreg %ptr) {
15981598
; GFX90A-NEXT: ; def v[0:3]
15991599
; GFX90A-NEXT: ;;#ASMEND
16001600
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
1601-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0] op_sel_hi:[0,0]
1601+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
16021602
; GFX90A-NEXT: global_store_dwordx2 v4, v[0:1], s[16:17]
16031603
; GFX90A-NEXT: s_waitcnt vmcnt(0)
16041604
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -1610,7 +1610,7 @@ define void @v_shuffle_v2f32_v4f32__3_2(ptr addrspace(1) inreg %ptr) {
16101610
; GFX940-NEXT: ; def v[0:3]
16111611
; GFX940-NEXT: ;;#ASMEND
16121612
; GFX940-NEXT: v_mov_b32_e32 v4, 0
1613-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0] op_sel_hi:[0,0]
1613+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[2:3] op_sel:[1,0]
16141614
; GFX940-NEXT: global_store_dwordx2 v4, v[0:1], s[0:1] sc0 sc1
16151615
; GFX940-NEXT: s_waitcnt vmcnt(0)
16161616
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -1687,7 +1687,7 @@ define void @v_shuffle_v2f32_v4f32__5_2(ptr addrspace(1) inreg %ptr) {
16871687
; GFX90A-NEXT: ;;#ASMSTART
16881688
; GFX90A-NEXT: ; def v[4:7]
16891689
; GFX90A-NEXT: ;;#ASMEND
1690-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[2:3] op_sel:[1,0] op_sel_hi:[0,0]
1690+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[2:3] op_sel:[1,0]
16911691
; GFX90A-NEXT: global_store_dwordx2 v8, v[0:1], s[16:17]
16921692
; GFX90A-NEXT: s_waitcnt vmcnt(0)
16931693
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -1703,7 +1703,7 @@ define void @v_shuffle_v2f32_v4f32__5_2(ptr addrspace(1) inreg %ptr) {
17031703
; GFX940-NEXT: ; def v[4:7]
17041704
; GFX940-NEXT: ;;#ASMEND
17051705
; GFX940-NEXT: s_nop 0
1706-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[2:3] op_sel:[1,0] op_sel_hi:[0,0]
1706+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[4:5], v[2:3] op_sel:[1,0]
17071707
; GFX940-NEXT: global_store_dwordx2 v8, v[0:1], s[0:1] sc0 sc1
17081708
; GFX940-NEXT: s_waitcnt vmcnt(0)
17091709
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -2319,7 +2319,7 @@ define void @v_shuffle_v2f32_v4f32__5_4(ptr addrspace(1) inreg %ptr) {
23192319
; GFX90A-NEXT: ; def v[0:3]
23202320
; GFX90A-NEXT: ;;#ASMEND
23212321
; GFX90A-NEXT: v_mov_b32_e32 v4, 0
2322-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
2322+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
23232323
; GFX90A-NEXT: global_store_dwordx2 v4, v[0:1], s[16:17]
23242324
; GFX90A-NEXT: s_waitcnt vmcnt(0)
23252325
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -2331,7 +2331,7 @@ define void @v_shuffle_v2f32_v4f32__5_4(ptr addrspace(1) inreg %ptr) {
23312331
; GFX940-NEXT: ; def v[0:3]
23322332
; GFX940-NEXT: ;;#ASMEND
23332333
; GFX940-NEXT: v_mov_b32_e32 v4, 0
2334-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0] op_sel_hi:[0,0]
2334+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[0:1] op_sel:[1,0]
23352335
; GFX940-NEXT: global_store_dwordx2 v4, v[0:1], s[0:1] sc0 sc1
23362336
; GFX940-NEXT: s_waitcnt vmcnt(0)
23372337
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -2884,7 +2884,7 @@ define void @v_shuffle_v2f32_v4f32__1_6(ptr addrspace(1) inreg %ptr) {
28842884
; GFX90A-NEXT: ;;#ASMSTART
28852885
; GFX90A-NEXT: ; def v[2:5]
28862886
; GFX90A-NEXT: ;;#ASMEND
2887-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[4:5] op_sel:[1,0] op_sel_hi:[0,0]
2887+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[4:5] op_sel:[1,0]
28882888
; GFX90A-NEXT: global_store_dwordx2 v6, v[0:1], s[16:17]
28892889
; GFX90A-NEXT: s_waitcnt vmcnt(0)
28902890
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -2900,7 +2900,7 @@ define void @v_shuffle_v2f32_v4f32__1_6(ptr addrspace(1) inreg %ptr) {
29002900
; GFX940-NEXT: ; def v[2:5]
29012901
; GFX940-NEXT: ;;#ASMEND
29022902
; GFX940-NEXT: s_nop 0
2903-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[4:5] op_sel:[1,0] op_sel_hi:[0,0]
2903+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[0:1], v[4:5] op_sel:[1,0]
29042904
; GFX940-NEXT: global_store_dwordx2 v6, v[0:1], s[0:1] sc0 sc1
29052905
; GFX940-NEXT: s_waitcnt vmcnt(0)
29062906
; GFX940-NEXT: s_setpc_b64 s[30:31]
@@ -2990,7 +2990,7 @@ define void @v_shuffle_v2f32_v4f32__3_6(ptr addrspace(1) inreg %ptr) {
29902990
; GFX90A-NEXT: ;;#ASMSTART
29912991
; GFX90A-NEXT: ; def v[4:7]
29922992
; GFX90A-NEXT: ;;#ASMEND
2993-
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[6:7] op_sel:[1,0] op_sel_hi:[0,0]
2993+
; GFX90A-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[6:7] op_sel:[1,0]
29942994
; GFX90A-NEXT: global_store_dwordx2 v8, v[0:1], s[16:17]
29952995
; GFX90A-NEXT: s_waitcnt vmcnt(0)
29962996
; GFX90A-NEXT: s_setpc_b64 s[30:31]
@@ -3006,7 +3006,7 @@ define void @v_shuffle_v2f32_v4f32__3_6(ptr addrspace(1) inreg %ptr) {
30063006
; GFX940-NEXT: ; def v[4:7]
30073007
; GFX940-NEXT: ;;#ASMEND
30083008
; GFX940-NEXT: s_nop 0
3009-
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[6:7] op_sel:[1,0] op_sel_hi:[0,0]
3009+
; GFX940-NEXT: v_pk_mov_b32 v[0:1], v[2:3], v[6:7] op_sel:[1,0]
30103010
; GFX940-NEXT: global_store_dwordx2 v8, v[0:1], s[0:1] sc0 sc1
30113011
; GFX940-NEXT: s_waitcnt vmcnt(0)
30123012
; GFX940-NEXT: s_setpc_b64 s[30:31]

0 commit comments

Comments
 (0)