Skip to content

Commit 054e7c5

Browse files
RKSimonRKSimon
committed
[VectorCombine] foldInsExtVectorToShuffle - ignore shuffle costs for 'identity' insertion masks
<u,1,u,u> 'inplace' single src shuffles can be treated as free identity shuffles - ignore any shuffle cost (similar to what we already do in other folds like foldShuffleOfShuffles) - eventually getShuffleCost should just return TCC_Free in these cases but in a lot of the targets' shuffle cost logic this currently ends up treated as a generic SK_PermuteSingleSrc. We still want to generate the shuffle as it will help further shuffle folds with the additional PoisonMaskElem 'undemanded' elements.
1 parent 327e2b7 commit 054e7c5

File tree

3 files changed

+22
-36
lines changed

3 files changed

+22
-36
lines changed

‎llvm/lib/Transforms/Vectorize/VectorCombine.cpp‎

Lines changed: 6 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -3097,8 +3097,12 @@ bool VectorCombine::foldInsExtVectorToShuffle(Instruction &I) {
30973097
TTI.getVectorInstrCost(*Ext, VecTy, CostKind, ExtIdx);
30983098
InstructionCost OldCost = ExtCost + InsCost;
30993099

3100-
InstructionCost NewCost = TTI.getShuffleCost(SK, VecTy, Mask, CostKind, 0,
3101-
nullptr, {DstVec, SrcVec});
3100+
// Ignore 'free' identity insertion shuffle.
3101+
// TODO: getShuffleCost should return TCC_Free for Identity shuffles.
3102+
InstructionCost NewCost = 0;
3103+
if (!ShuffleVectorInst::isIdentityMask(Mask, NumElts))
3104+
NewCost += TTI.getShuffleCost(SK, VecTy, Mask, CostKind, 0, nullptr,
3105+
{DstVec, SrcVec});
31023106
if (!Ext->hasOneUse())
31033107
NewCost += ExtCost;
31043108

‎llvm/test/Transforms/PhaseOrdering/X86/hadd.ll‎

Lines changed: 15 additions & 30 deletions
Original file line numberDiff line numberDiff line change
@@ -512,13 +512,10 @@ define <4 x float> @add_v4f32_u123(<4 x float> %a, <4 x float> %b) {
512512

513513
define <4 x float> @add_v4f32_0u23(<4 x float> %a, <4 x float> %b) {
514514
; CHECK-LABEL: @add_v4f32_0u23(
515-
; CHECK-NEXT: [[SHIFT:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> poison, <4 x i32> <i32 1, i32 poison, i32 poison, i32 poison>
516-
; CHECK-NEXT: [[TMP1:%.*]] = fadd <4 x float> [[A]], [[SHIFT]]
517-
; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[B:%.*]], <4 x float> poison, <4 x i32> <i32 1, i32 2, i32 poison, i32 poison>
518-
; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[B]], <4 x float> poison, <4 x i32> <i32 0, i32 3, i32 poison, i32 poison>
515+
; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> [[B:%.*]], <4 x i32> <i32 0, i32 poison, i32 5, i32 6>
516+
; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B]], <4 x i32> <i32 1, i32 poison, i32 4, i32 7>
519517
; CHECK-NEXT: [[TMP4:%.*]] = fadd <4 x float> [[TMP2]], [[TMP3]]
520-
; CHECK-NEXT: [[RESULT1:%.*]] = shufflevector <4 x float> [[TMP1]], <4 x float> [[TMP4]], <4 x i32> <i32 0, i32 poison, i32 4, i32 5>
521-
; CHECK-NEXT: ret <4 x float> [[RESULT1]]
518+
; CHECK-NEXT: ret <4 x float> [[TMP4]]
522519
;
523520
%a0 = extractelement <4 x float> %a, i32 0
524521
%a1 = extractelement <4 x float> %a, i32 1
@@ -542,13 +539,10 @@ define <4 x float> @add_v4f32_0u23(<4 x float> %a, <4 x float> %b) {
542539

543540
define <4 x float> @add_v4f32_01u3(<4 x float> %a, <4 x float> %b) {
544541
; SSE2-LABEL: @add_v4f32_01u3(
545-
; SSE2-NEXT: [[SHIFT:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> poison, <4 x i32> <i32 1, i32 poison, i32 poison, i32 poison>
546-
; SSE2-NEXT: [[TMP1:%.*]] = fadd <4 x float> [[A]], [[SHIFT]]
547-
; SSE2-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B:%.*]], <4 x i32> <i32 2, i32 poison, i32 6, i32 poison>
548-
; SSE2-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B]], <4 x i32> <i32 3, i32 poison, i32 7, i32 poison>
542+
; SSE2-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> [[B:%.*]], <4 x i32> <i32 0, i32 2, i32 poison, i32 6>
543+
; SSE2-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B]], <4 x i32> <i32 1, i32 3, i32 poison, i32 7>
549544
; SSE2-NEXT: [[TMP4:%.*]] = fadd <4 x float> [[TMP2]], [[TMP3]]
550-
; SSE2-NEXT: [[RESULT1:%.*]] = shufflevector <4 x float> [[TMP1]], <4 x float> [[TMP4]], <4 x i32> <i32 0, i32 4, i32 poison, i32 6>
551-
; SSE2-NEXT: ret <4 x float> [[RESULT1]]
545+
; SSE2-NEXT: ret <4 x float> [[TMP4]]
552546
;
553547
; SSE4-LABEL: @add_v4f32_01u3(
554548
; SSE4-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> [[B:%.*]], <4 x i32> <i32 1, i32 2, i32 poison, i32 6>
@@ -563,13 +557,10 @@ define <4 x float> @add_v4f32_01u3(<4 x float> %a, <4 x float> %b) {
563557
; AVX2-NEXT: ret <4 x float> [[TMP4]]
564558
;
565559
; AVX512-LABEL: @add_v4f32_01u3(
566-
; AVX512-NEXT: [[SHIFT:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> poison, <4 x i32> <i32 1, i32 poison, i32 poison, i32 poison>
567-
; AVX512-NEXT: [[TMP1:%.*]] = fadd <4 x float> [[A]], [[SHIFT]]
568-
; AVX512-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B:%.*]], <4 x i32> <i32 2, i32 poison, i32 6, i32 poison>
569-
; AVX512-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B]], <4 x i32> <i32 3, i32 poison, i32 7, i32 poison>
560+
; AVX512-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> [[B:%.*]], <4 x i32> <i32 0, i32 2, i32 poison, i32 6>
561+
; AVX512-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B]], <4 x i32> <i32 1, i32 3, i32 poison, i32 7>
570562
; AVX512-NEXT: [[TMP4:%.*]] = fadd <4 x float> [[TMP2]], [[TMP3]]
571-
; AVX512-NEXT: [[RESULT1:%.*]] = shufflevector <4 x float> [[TMP1]], <4 x float> [[TMP4]], <4 x i32> <i32 0, i32 4, i32 poison, i32 6>
572-
; AVX512-NEXT: ret <4 x float> [[RESULT1]]
563+
; AVX512-NEXT: ret <4 x float> [[TMP4]]
573564
;
574565
%a0 = extractelement <4 x float> %a, i32 0
575566
%a1 = extractelement <4 x float> %a, i32 1
@@ -593,13 +584,10 @@ define <4 x float> @add_v4f32_01u3(<4 x float> %a, <4 x float> %b) {
593584

594585
define <4 x float> @add_v4f32_012u(<4 x float> %a, <4 x float> %b) {
595586
; SSE2-LABEL: @add_v4f32_012u(
596-
; SSE2-NEXT: [[SHIFT:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> poison, <4 x i32> <i32 1, i32 poison, i32 poison, i32 poison>
597-
; SSE2-NEXT: [[TMP1:%.*]] = fadd <4 x float> [[A]], [[SHIFT]]
598-
; SSE2-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B:%.*]], <4 x i32> <i32 2, i32 4, i32 poison, i32 poison>
599-
; SSE2-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B]], <4 x i32> <i32 3, i32 5, i32 poison, i32 poison>
587+
; SSE2-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> [[B:%.*]], <4 x i32> <i32 0, i32 2, i32 4, i32 poison>
588+
; SSE2-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B]], <4 x i32> <i32 1, i32 3, i32 5, i32 poison>
600589
; SSE2-NEXT: [[TMP4:%.*]] = fadd <4 x float> [[TMP2]], [[TMP3]]
601-
; SSE2-NEXT: [[RESULT1:%.*]] = shufflevector <4 x float> [[TMP1]], <4 x float> [[TMP4]], <4 x i32> <i32 0, i32 4, i32 5, i32 poison>
602-
; SSE2-NEXT: ret <4 x float> [[RESULT1]]
590+
; SSE2-NEXT: ret <4 x float> [[TMP4]]
603591
;
604592
; SSE4-LABEL: @add_v4f32_012u(
605593
; SSE4-NEXT: [[SHIFT:%.*]] = shufflevector <4 x float> [[B:%.*]], <4 x float> poison, <4 x i32> <i32 1, i32 poison, i32 poison, i32 poison>
@@ -620,13 +608,10 @@ define <4 x float> @add_v4f32_012u(<4 x float> %a, <4 x float> %b) {
620608
; AVX2-NEXT: ret <4 x float> [[RESULT]]
621609
;
622610
; AVX512-LABEL: @add_v4f32_012u(
623-
; AVX512-NEXT: [[SHIFT:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> poison, <4 x i32> <i32 1, i32 poison, i32 poison, i32 poison>
624-
; AVX512-NEXT: [[TMP1:%.*]] = fadd <4 x float> [[A]], [[SHIFT]]
625-
; AVX512-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B:%.*]], <4 x i32> <i32 2, i32 4, i32 poison, i32 poison>
626-
; AVX512-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B]], <4 x i32> <i32 3, i32 5, i32 poison, i32 poison>
611+
; AVX512-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> [[B:%.*]], <4 x i32> <i32 0, i32 2, i32 4, i32 poison>
612+
; AVX512-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[A]], <4 x float> [[B]], <4 x i32> <i32 1, i32 3, i32 5, i32 poison>
627613
; AVX512-NEXT: [[TMP4:%.*]] = fadd <4 x float> [[TMP2]], [[TMP3]]
628-
; AVX512-NEXT: [[RESULT1:%.*]] = shufflevector <4 x float> [[TMP1]], <4 x float> [[TMP4]], <4 x i32> <i32 0, i32 4, i32 5, i32 poison>
629-
; AVX512-NEXT: ret <4 x float> [[RESULT1]]
614+
; AVX512-NEXT: ret <4 x float> [[TMP4]]
630615
;
631616
%a0 = extractelement <4 x float> %a, i32 0
632617
%a1 = extractelement <4 x float> %a, i32 1

‎llvm/test/Transforms/VectorCombine/X86/load-inseltpoison.ll‎

Lines changed: 1 addition & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -544,10 +544,7 @@ define void @PR47558_multiple_use_load(ptr nocapture nonnull %resultptr, ptr noc
544544
; CHECK-NEXT: [[T1:%.*]] = insertelement <2 x float> poison, float [[SCALE]], i32 0
545545
; CHECK-NEXT: [[T2:%.*]] = insertelement <2 x float> [[T1]], float [[SCALE]], i32 1
546546
; CHECK-NEXT: [[T3:%.*]] = fmul <2 x float> [[OP]], [[T2]]
547-
; CHECK-NEXT: [[T4:%.*]] = extractelement <2 x float> [[T3]], i32 0
548-
; CHECK-NEXT: [[RESULT0:%.*]] = insertelement <2 x float> poison, float [[T4]], i32 0
549-
; CHECK-NEXT: [[RESULT1:%.*]] = shufflevector <2 x float> [[RESULT0]], <2 x float> [[T3]], <2 x i32> <i32 0, i32 3>
550-
; CHECK-NEXT: store <2 x float> [[RESULT1]], ptr [[RESULTPTR:%.*]], align 8
547+
; CHECK-NEXT: store <2 x float> [[T3]], ptr [[RESULTPTR:%.*]], align 8
551548
; CHECK-NEXT: ret void
552549
;
553550
%scaleptr = tail call nonnull align 16 dereferenceable(64) ptr @getscaleptr()

0 commit comments

Comments
 (0)