[llvm] 50ba89a - [VectorCombine] support mismatching extract/insert indices for foldInsExtFNeg (#126408)

via llvm-commits llvm-commits at lists.llvm.org
Fri Nov 7 10:35:44 PST 2025


Author: hanbeom
Date: 2025-11-07T18:35:40Z
New Revision: 50ba89a22e452b47f79112e7277f40cbdab27bdf

URL: https://github.com/llvm/llvm-project/commit/50ba89a22e452b47f79112e7277f40cbdab27bdf
DIFF: https://github.com/llvm/llvm-project/commit/50ba89a22e452b47f79112e7277f40cbdab27bdf.diff

LOG: [VectorCombine] support mismatching extract/insert indices for foldInsExtFNeg (#126408)

insertelt DestVec, (fneg (extractelt SrcVec, Index)), Index 
-> shuffle DestVec, (shuffle (fneg SrcVec), poison, SrcMask), Mask

In previous, the above transform was only possible if the Extract/Insert
Index was the same; this patch makes the above transform possible
even if the two indexes are different.

Proof: https://alive2.llvm.org/ce/z/aDfdyG
Fixes: https://github.com/llvm/llvm-project/issues/125675

Added: 
    

Modified: 
    llvm/lib/Transforms/Vectorize/VectorCombine.cpp
    llvm/test/Transforms/PhaseOrdering/X86/addsub-inseltpoison.ll
    llvm/test/Transforms/PhaseOrdering/X86/addsub.ll
    llvm/test/Transforms/VectorCombine/X86/extract-fneg-insert.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Vectorize/VectorCombine.cpp b/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
index 27a8bbd5776be..ed3a0a0ab023d 100644
--- a/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
+++ b/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
@@ -696,11 +696,11 @@ bool VectorCombine::foldExtractExtract(Instruction &I) {
 /// shuffle.
 bool VectorCombine::foldInsExtFNeg(Instruction &I) {
   // Match an insert (op (extract)) pattern.
-  Value *DestVec;
-  uint64_t Index;
+  Value *DstVec;
+  uint64_t ExtIdx, InsIdx;
   Instruction *FNeg;
-  if (!match(&I, m_InsertElt(m_Value(DestVec), m_OneUse(m_Instruction(FNeg)),
-                             m_ConstantInt(Index))))
+  if (!match(&I, m_InsertElt(m_Value(DstVec), m_OneUse(m_Instruction(FNeg)),
+                             m_ConstantInt(InsIdx))))
     return false;
 
   // Note: This handles the canonical fneg instruction and "fsub -0.0, X".
@@ -708,67 +708,74 @@ bool VectorCombine::foldInsExtFNeg(Instruction &I) {
   Instruction *Extract;
   if (!match(FNeg, m_FNeg(m_CombineAnd(
                        m_Instruction(Extract),
-                       m_ExtractElt(m_Value(SrcVec), m_SpecificInt(Index))))))
+                       m_ExtractElt(m_Value(SrcVec), m_ConstantInt(ExtIdx))))))
     return false;
 
-  auto *VecTy = cast<FixedVectorType>(I.getType());
-  auto *ScalarTy = VecTy->getScalarType();
+  auto *DstVecTy = cast<FixedVectorType>(DstVec->getType());
+  auto *DstVecScalarTy = DstVecTy->getScalarType();
   auto *SrcVecTy = dyn_cast<FixedVectorType>(SrcVec->getType());
-  if (!SrcVecTy || ScalarTy != SrcVecTy->getScalarType())
+  if (!SrcVecTy || DstVecScalarTy != SrcVecTy->getScalarType())
     return false;
 
-  // Ignore bogus insert/extract index.
-  unsigned NumElts = VecTy->getNumElements();
-  if (Index >= NumElts)
+  // Ignore if insert/extract index is out of bounds or destination vector has
+  // one element
+  unsigned NumDstElts = DstVecTy->getNumElements();
+  unsigned NumSrcElts = SrcVecTy->getNumElements();
+  if (ExtIdx > NumSrcElts || InsIdx >= NumDstElts || NumDstElts == 1)
     return false;
 
   // We are inserting the negated element into the same lane that we extracted
   // from. This is equivalent to a select-shuffle that chooses all but the
   // negated element from the destination vector.
-  SmallVector<int> Mask(NumElts);
+  SmallVector<int> Mask(NumDstElts);
   std::iota(Mask.begin(), Mask.end(), 0);
-  Mask[Index] = Index + NumElts;
+  Mask[InsIdx] = (ExtIdx % NumDstElts) + NumDstElts;
   InstructionCost OldCost =
-      TTI.getArithmeticInstrCost(Instruction::FNeg, ScalarTy, CostKind) +
-      TTI.getVectorInstrCost(I, VecTy, CostKind, Index);
+      TTI.getArithmeticInstrCost(Instruction::FNeg, DstVecScalarTy, CostKind) +
+      TTI.getVectorInstrCost(I, DstVecTy, CostKind, InsIdx);
 
   // If the extract has one use, it will be eliminated, so count it in the
   // original cost. If it has more than one use, ignore the cost because it will
   // be the same before/after.
   if (Extract->hasOneUse())
-    OldCost += TTI.getVectorInstrCost(*Extract, VecTy, CostKind, Index);
+    OldCost += TTI.getVectorInstrCost(*Extract, SrcVecTy, CostKind, ExtIdx);
 
   InstructionCost NewCost =
-      TTI.getArithmeticInstrCost(Instruction::FNeg, VecTy, CostKind) +
-      TTI.getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, VecTy, VecTy,
-                         Mask, CostKind);
+      TTI.getArithmeticInstrCost(Instruction::FNeg, SrcVecTy, CostKind) +
+      TTI.getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, DstVecTy,
+                         DstVecTy, Mask, CostKind);
 
-  bool NeedLenChg = SrcVecTy->getNumElements() != NumElts;
+  bool NeedLenChg = SrcVecTy->getNumElements() != NumDstElts;
   // If the lengths of the two vectors are not equal,
   // we need to add a length-change vector. Add this cost.
   SmallVector<int> SrcMask;
   if (NeedLenChg) {
-    SrcMask.assign(NumElts, PoisonMaskElem);
-    SrcMask[Index] = Index;
+    SrcMask.assign(NumDstElts, PoisonMaskElem);
+    SrcMask[ExtIdx % NumDstElts] = ExtIdx;
     NewCost += TTI.getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc,
-                                  VecTy, SrcVecTy, SrcMask, CostKind);
+                                  DstVecTy, SrcVecTy, SrcMask, CostKind);
   }
 
+  LLVM_DEBUG(dbgs() << "Found an insertion of (extract)fneg : " << I
+                    << "\n  OldCost: " << OldCost << " vs NewCost: " << NewCost
+                    << "\n");
   if (NewCost > OldCost)
     return false;
 
-  Value *NewShuf;
-  // insertelt DestVec, (fneg (extractelt SrcVec, Index)), Index
+  Value *NewShuf, *LenChgShuf = nullptr;
+  // insertelt DstVec, (fneg (extractelt SrcVec, Index)), Index
   Value *VecFNeg = Builder.CreateFNegFMF(SrcVec, FNeg);
   if (NeedLenChg) {
-    // shuffle DestVec, (shuffle (fneg SrcVec), poison, SrcMask), Mask
-    Value *LenChgShuf = Builder.CreateShuffleVector(VecFNeg, SrcMask);
-    NewShuf = Builder.CreateShuffleVector(DestVec, LenChgShuf, Mask);
+    // shuffle DstVec, (shuffle (fneg SrcVec), poison, SrcMask), Mask
+    LenChgShuf = Builder.CreateShuffleVector(VecFNeg, SrcMask);
+    NewShuf = Builder.CreateShuffleVector(DstVec, LenChgShuf, Mask);
+    Worklist.pushValue(LenChgShuf);
   } else {
-    // shuffle DestVec, (fneg SrcVec), Mask
-    NewShuf = Builder.CreateShuffleVector(DestVec, VecFNeg, Mask);
+    // shuffle DstVec, (fneg SrcVec), Mask
+    NewShuf = Builder.CreateShuffleVector(DstVec, VecFNeg, Mask);
   }
 
+  Worklist.pushValue(VecFNeg);
   replaceValue(I, *NewShuf);
   return true;
 }

diff  --git a/llvm/test/Transforms/PhaseOrdering/X86/addsub-inseltpoison.ll b/llvm/test/Transforms/PhaseOrdering/X86/addsub-inseltpoison.ll
index 2c1d73eaafc5e..9f3244ded92ff 100644
--- a/llvm/test/Transforms/PhaseOrdering/X86/addsub-inseltpoison.ll
+++ b/llvm/test/Transforms/PhaseOrdering/X86/addsub-inseltpoison.ll
@@ -498,11 +498,9 @@ define void @add_aggregate_store(<2 x float> %a0, <2 x float> %a1, <2 x float> %
 ; PR58139
 define <2 x double> @_mm_complexmult_pd_naive(<2 x double> %a, <2 x double> %b) {
 ; SSE-LABEL: @_mm_complexmult_pd_naive(
-; SSE-NEXT:    [[B1:%.*]] = extractelement <2 x double> [[B:%.*]], i64 1
-; SSE-NEXT:    [[TMP1:%.*]] = fneg double [[B1]]
 ; SSE-NEXT:    [[TMP2:%.*]] = shufflevector <2 x double> [[A:%.*]], <2 x double> poison, <2 x i32> <i32 1, i32 1>
-; SSE-NEXT:    [[TMP3:%.*]] = shufflevector <2 x double> [[B]], <2 x double> poison, <2 x i32> <i32 poison, i32 0>
-; SSE-NEXT:    [[TMP4:%.*]] = insertelement <2 x double> [[TMP3]], double [[TMP1]], i64 0
+; SSE-NEXT:    [[TMP3:%.*]] = fneg <2 x double> [[B:%.*]]
+; SSE-NEXT:    [[TMP4:%.*]] = shufflevector <2 x double> [[TMP3]], <2 x double> [[B]], <2 x i32> <i32 1, i32 2>
 ; SSE-NEXT:    [[TMP5:%.*]] = fmul <2 x double> [[TMP2]], [[TMP4]]
 ; SSE-NEXT:    [[TMP6:%.*]] = shufflevector <2 x double> [[A]], <2 x double> poison, <2 x i32> zeroinitializer
 ; SSE-NEXT:    [[TMP7:%.*]] = tail call <2 x double> @llvm.fmuladd.v2f64(<2 x double> [[TMP6]], <2 x double> [[B]], <2 x double> [[TMP5]])

diff  --git a/llvm/test/Transforms/PhaseOrdering/X86/addsub.ll b/llvm/test/Transforms/PhaseOrdering/X86/addsub.ll
index fa6403f3d4267..de64bf2657f72 100644
--- a/llvm/test/Transforms/PhaseOrdering/X86/addsub.ll
+++ b/llvm/test/Transforms/PhaseOrdering/X86/addsub.ll
@@ -502,11 +502,9 @@ define void @add_aggregate_store(<2 x float> %a0, <2 x float> %a1, <2 x float> %
 ; PR58139
 define <2 x double> @_mm_complexmult_pd_naive(<2 x double> %a, <2 x double> %b) {
 ; SSE-LABEL: @_mm_complexmult_pd_naive(
-; SSE-NEXT:    [[B1:%.*]] = extractelement <2 x double> [[B:%.*]], i64 1
-; SSE-NEXT:    [[TMP1:%.*]] = fneg double [[B1]]
 ; SSE-NEXT:    [[TMP2:%.*]] = shufflevector <2 x double> [[A:%.*]], <2 x double> poison, <2 x i32> <i32 1, i32 1>
-; SSE-NEXT:    [[TMP3:%.*]] = shufflevector <2 x double> [[B]], <2 x double> poison, <2 x i32> <i32 poison, i32 0>
-; SSE-NEXT:    [[TMP4:%.*]] = insertelement <2 x double> [[TMP3]], double [[TMP1]], i64 0
+; SSE-NEXT:    [[TMP3:%.*]] = fneg <2 x double> [[B:%.*]]
+; SSE-NEXT:    [[TMP4:%.*]] = shufflevector <2 x double> [[TMP3]], <2 x double> [[B]], <2 x i32> <i32 1, i32 2>
 ; SSE-NEXT:    [[TMP5:%.*]] = fmul <2 x double> [[TMP2]], [[TMP4]]
 ; SSE-NEXT:    [[TMP6:%.*]] = shufflevector <2 x double> [[A]], <2 x double> poison, <2 x i32> zeroinitializer
 ; SSE-NEXT:    [[TMP7:%.*]] = tail call <2 x double> @llvm.fmuladd.v2f64(<2 x double> [[TMP6]], <2 x double> [[B]], <2 x double> [[TMP5]])

diff  --git a/llvm/test/Transforms/VectorCombine/X86/extract-fneg-insert.ll b/llvm/test/Transforms/VectorCombine/X86/extract-fneg-insert.ll
index 5358e0419e7a7..88fcf359f7c8e 100644
--- a/llvm/test/Transforms/VectorCombine/X86/extract-fneg-insert.ll
+++ b/llvm/test/Transforms/VectorCombine/X86/extract-fneg-insert.ll
@@ -58,6 +58,19 @@ define <4 x float> @ext2_v2f32v4f32(<2 x float> %x, <4 x float> %y) {
   ret <4 x float> %r
 }
 
+define <2 x float> @ext2_v4f32v2f32(<4 x float> %x, <2 x float> %y) {
+; CHECK-LABEL: @ext2_v4f32v2f32(
+; CHECK-NEXT:    [[TMP1:%.*]] = fneg <4 x float> [[X:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <4 x float> [[TMP1]], <4 x float> poison, <2 x i32> <i32 poison, i32 3>
+; CHECK-NEXT:    [[R:%.*]] = shufflevector <2 x float> [[Y:%.*]], <2 x float> [[TMP2]], <2 x i32> <i32 0, i32 3>
+; CHECK-NEXT:    ret <2 x float> [[R]]
+;
+  %e = extractelement <4 x float> %x, i32 3
+  %n = fneg float %e
+  %r = insertelement <2 x float> %y, float %n, i32 1
+  ret <2 x float> %r
+}
+
 ; Eliminating extract/insert is still profitable. Flags propagate.
 
 define <2 x double> @ext1_v2f64(<2 x double> %x, <2 x double> %y) {
@@ -73,17 +86,11 @@ define <2 x double> @ext1_v2f64(<2 x double> %x, <2 x double> %y) {
 }
 
 define <4 x double> @ext1_v2f64v4f64(<2 x double> %x, <4 x double> %y) {
-; SSE-LABEL: @ext1_v2f64v4f64(
-; SSE-NEXT:    [[E:%.*]] = extractelement <2 x double> [[X:%.*]], i32 1
-; SSE-NEXT:    [[N:%.*]] = fneg nsz double [[E]]
-; SSE-NEXT:    [[R:%.*]] = insertelement <4 x double> [[Y:%.*]], double [[N]], i32 1
-; SSE-NEXT:    ret <4 x double> [[R]]
-;
-; AVX-LABEL: @ext1_v2f64v4f64(
-; AVX-NEXT:    [[TMP1:%.*]] = fneg nsz <2 x double> [[X:%.*]]
-; AVX-NEXT:    [[TMP2:%.*]] = shufflevector <2 x double> [[TMP1]], <2 x double> poison, <4 x i32> <i32 poison, i32 1, i32 poison, i32 poison>
-; AVX-NEXT:    [[R:%.*]] = shufflevector <4 x double> [[Y:%.*]], <4 x double> [[TMP2]], <4 x i32> <i32 0, i32 5, i32 2, i32 3>
-; AVX-NEXT:    ret <4 x double> [[R]]
+; CHECK-LABEL: @ext1_v2f64v4f64(
+; CHECK-NEXT:    [[TMP1:%.*]] = fneg nsz <2 x double> [[X:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <2 x double> [[TMP1]], <2 x double> poison, <4 x i32> <i32 poison, i32 1, i32 poison, i32 poison>
+; CHECK-NEXT:    [[R:%.*]] = shufflevector <4 x double> [[Y:%.*]], <4 x double> [[TMP2]], <4 x i32> <i32 0, i32 5, i32 2, i32 3>
+; CHECK-NEXT:    ret <4 x double> [[R]]
 ;
   %e = extractelement <2 x double> %x, i32 1
   %n = fneg nsz double %e
@@ -91,6 +98,19 @@ define <4 x double> @ext1_v2f64v4f64(<2 x double> %x, <4 x double> %y) {
   ret <4 x double> %r
 }
 
+define <2 x double> @ext1_v4f64v2f64(<4 x double> %x, <2 x double> %y) {
+; CHECK-LABEL: @ext1_v4f64v2f64(
+; CHECK-NEXT:    [[TMP1:%.*]] = fneg nsz <4 x double> [[X:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <4 x double> [[TMP1]], <4 x double> poison, <2 x i32> <i32 poison, i32 3>
+; CHECK-NEXT:    [[R:%.*]] = shufflevector <2 x double> [[Y:%.*]], <2 x double> [[TMP2]], <2 x i32> <i32 0, i32 3>
+; CHECK-NEXT:    ret <2 x double> [[R]]
+;
+  %e = extractelement <4 x double> %x, i32 3
+  %n = fneg nsz double %e
+  %r = insertelement <2 x double> %y, double %n, i32 1
+  ret <2 x double> %r
+}
+
 define <8 x float> @ext7_v8f32(<8 x float> %x, <8 x float> %y) {
 ; CHECK-LABEL: @ext7_v8f32(
 ; CHECK-NEXT:    [[TMP1:%.*]] = fneg <8 x float> [[X:%.*]]
@@ -105,9 +125,9 @@ define <8 x float> @ext7_v8f32(<8 x float> %x, <8 x float> %y) {
 
 define <8 x float> @ext7_v4f32v8f32(<4 x float> %x, <8 x float> %y) {
 ; CHECK-LABEL: @ext7_v4f32v8f32(
-; CHECK-NEXT:    [[E:%.*]] = extractelement <4 x float> [[X:%.*]], i32 3
-; CHECK-NEXT:    [[N:%.*]] = fneg float [[E]]
-; CHECK-NEXT:    [[R:%.*]] = insertelement <8 x float> [[Y:%.*]], float [[N]], i32 7
+; CHECK-NEXT:    [[TMP1:%.*]] = fneg <4 x float> [[X:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <4 x float> [[TMP1]], <4 x float> poison, <8 x i32> <i32 poison, i32 poison, i32 poison, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    [[R:%.*]] = shufflevector <8 x float> [[Y:%.*]], <8 x float> [[TMP2]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 11>
 ; CHECK-NEXT:    ret <8 x float> [[R]]
 ;
   %e = extractelement <4 x float> %x, i32 3
@@ -116,6 +136,19 @@ define <8 x float> @ext7_v4f32v8f32(<4 x float> %x, <8 x float> %y) {
   ret <8 x float> %r
 }
 
+define <4 x float> @ext7_v8f32v4f32(<8 x float> %x, <4 x float> %y) {
+; CHECK-LABEL: @ext7_v8f32v4f32(
+; CHECK-NEXT:    [[TMP1:%.*]] = fneg <8 x float> [[X:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <8 x float> [[TMP1]], <8 x float> poison, <4 x i32> <i32 poison, i32 poison, i32 poison, i32 7>
+; CHECK-NEXT:    [[R:%.*]] = shufflevector <4 x float> [[Y:%.*]], <4 x float> [[TMP2]], <4 x i32> <i32 0, i32 1, i32 2, i32 7>
+; CHECK-NEXT:    ret <4 x float> [[R]]
+;
+  %e = extractelement <8 x float> %x, i32 7
+  %n = fneg float %e
+  %r = insertelement <4 x float> %y, float %n, i32 3
+  ret <4 x float> %r
+}
+
 ; Same as above with an extra use of the extracted element.
 
 define <8 x float> @ext7_v8f32_use1(<8 x float> %x, <8 x float> %y) {
@@ -141,12 +174,20 @@ define <8 x float> @ext7_v8f32_use1(<8 x float> %x, <8 x float> %y) {
 }
 
 define <8 x float> @ext7_v4f32v8f32_use1(<4 x float> %x, <8 x float> %y) {
-; CHECK-LABEL: @ext7_v4f32v8f32_use1(
-; CHECK-NEXT:    [[E:%.*]] = extractelement <4 x float> [[X:%.*]], i32 3
-; CHECK-NEXT:    call void @use(float [[E]])
-; CHECK-NEXT:    [[N:%.*]] = fneg float [[E]]
-; CHECK-NEXT:    [[R:%.*]] = insertelement <8 x float> [[Y:%.*]], float [[N]], i32 3
-; CHECK-NEXT:    ret <8 x float> [[R]]
+; SSE-LABEL: @ext7_v4f32v8f32_use1(
+; SSE-NEXT:    [[E:%.*]] = extractelement <4 x float> [[X:%.*]], i32 3
+; SSE-NEXT:    call void @use(float [[E]])
+; SSE-NEXT:    [[TMP1:%.*]] = fneg <4 x float> [[X]]
+; SSE-NEXT:    [[TMP2:%.*]] = shufflevector <4 x float> [[TMP1]], <4 x float> poison, <8 x i32> <i32 poison, i32 poison, i32 poison, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>
+; SSE-NEXT:    [[R:%.*]] = shufflevector <8 x float> [[Y:%.*]], <8 x float> [[TMP2]], <8 x i32> <i32 0, i32 1, i32 2, i32 11, i32 4, i32 5, i32 6, i32 7>
+; SSE-NEXT:    ret <8 x float> [[R]]
+;
+; AVX-LABEL: @ext7_v4f32v8f32_use1(
+; AVX-NEXT:    [[E:%.*]] = extractelement <4 x float> [[X:%.*]], i32 3
+; AVX-NEXT:    call void @use(float [[E]])
+; AVX-NEXT:    [[N:%.*]] = fneg float [[E]]
+; AVX-NEXT:    [[R:%.*]] = insertelement <8 x float> [[Y:%.*]], float [[N]], i32 3
+; AVX-NEXT:    ret <8 x float> [[R]]
 ;
   %e = extractelement <4 x float> %x, i32 3
   call void @use(float %e)
@@ -155,6 +196,29 @@ define <8 x float> @ext7_v4f32v8f32_use1(<4 x float> %x, <8 x float> %y) {
   ret <8 x float> %r
 }
 
+define <4 x float> @ext7_v8f32v4f32_use1(<8 x float> %x, <4 x float> %y) {
+; SSE-LABEL: @ext7_v8f32v4f32_use1(
+; SSE-NEXT:    [[E:%.*]] = extractelement <8 x float> [[X:%.*]], i32 7
+; SSE-NEXT:    call void @use(float [[E]])
+; SSE-NEXT:    [[TMP1:%.*]] = fneg <8 x float> [[X]]
+; SSE-NEXT:    [[TMP2:%.*]] = shufflevector <8 x float> [[TMP1]], <8 x float> poison, <4 x i32> <i32 poison, i32 poison, i32 poison, i32 7>
+; SSE-NEXT:    [[R:%.*]] = shufflevector <4 x float> [[Y:%.*]], <4 x float> [[TMP2]], <4 x i32> <i32 0, i32 1, i32 2, i32 7>
+; SSE-NEXT:    ret <4 x float> [[R]]
+;
+; AVX-LABEL: @ext7_v8f32v4f32_use1(
+; AVX-NEXT:    [[E:%.*]] = extractelement <8 x float> [[X:%.*]], i32 7
+; AVX-NEXT:    call void @use(float [[E]])
+; AVX-NEXT:    [[N:%.*]] = fneg float [[E]]
+; AVX-NEXT:    [[R:%.*]] = insertelement <4 x float> [[Y:%.*]], float [[N]], i32 3
+; AVX-NEXT:    ret <4 x float> [[R]]
+;
+  %e = extractelement <8 x float> %x, i32 7
+  call void @use(float %e)
+  %n = fneg float %e
+  %r = insertelement <4 x float> %y, float %n, i32 3
+  ret <4 x float> %r
+}
+
 ; Negative test - the transform is likely not profitable if the fneg has another use.
 
 define <8 x float> @ext7_v8f32_use2(<8 x float> %x, <8 x float> %y) {
@@ -187,6 +251,21 @@ define <8 x float> @ext7_v4f32v8f32_use2(<4 x float> %x, <8 x float> %y) {
   ret <8 x float> %r
 }
 
+define <4 x float> @ext7_v8f32v4f32_use2(<8 x float> %x, <4 x float> %y) {
+; CHECK-LABEL: @ext7_v8f32v4f32_use2(
+; CHECK-NEXT:    [[E:%.*]] = extractelement <8 x float> [[X:%.*]], i32 7
+; CHECK-NEXT:    [[N:%.*]] = fneg float [[E]]
+; CHECK-NEXT:    call void @use(float [[N]])
+; CHECK-NEXT:    [[R:%.*]] = insertelement <4 x float> [[Y:%.*]], float [[N]], i32 3
+; CHECK-NEXT:    ret <4 x float> [[R]]
+;
+  %e = extractelement <8 x float> %x, i32 7
+  %n = fneg float %e
+  call void @use(float %n)
+  %r = insertelement <4 x float> %y, float %n, i32 3
+  ret <4 x float> %r
+}
+
 ; Negative test - can't convert variable index to a shuffle.
 
 define <2 x double> @ext_index_var_v2f64(<2 x double> %x, <2 x double> %y, i32 %index) {
@@ -215,14 +294,10 @@ define <4 x double> @ext_index_var_v2f64v4f64(<2 x double> %x, <4 x double> %y,
   ret <4 x double> %r
 }
 
-; Negative test - require same extract/insert index for simple shuffle.
-; TODO: We could handle this by adjusting the cost calculation.
-
 define <2 x double> @ext1_v2f64_ins0(<2 x double> %x, <2 x double> %y) {
 ; CHECK-LABEL: @ext1_v2f64_ins0(
-; CHECK-NEXT:    [[E:%.*]] = extractelement <2 x double> [[X:%.*]], i32 1
-; CHECK-NEXT:    [[N:%.*]] = fneg nsz double [[E]]
-; CHECK-NEXT:    [[R:%.*]] = insertelement <2 x double> [[Y:%.*]], double [[N]], i32 0
+; CHECK-NEXT:    [[TMP1:%.*]] = fneg nsz <2 x double> [[X:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = shufflevector <2 x double> [[Y:%.*]], <2 x double> [[TMP1]], <2 x i32> <i32 3, i32 1>
 ; CHECK-NEXT:    ret <2 x double> [[R]]
 ;
   %e = extractelement <2 x double> %x, i32 1
@@ -231,12 +306,11 @@ define <2 x double> @ext1_v2f64_ins0(<2 x double> %x, <2 x double> %y) {
   ret <2 x double> %r
 }
 
-; Negative test - extract from an index greater than the vector width of the destination
 define <2 x double> @ext3_v4f64v2f64(<4 x double> %x, <2 x double> %y) {
 ; CHECK-LABEL: @ext3_v4f64v2f64(
-; CHECK-NEXT:    [[E:%.*]] = extractelement <4 x double> [[X:%.*]], i32 3
-; CHECK-NEXT:    [[N:%.*]] = fneg nsz double [[E]]
-; CHECK-NEXT:    [[R:%.*]] = insertelement <2 x double> [[Y:%.*]], double [[N]], i32 1
+; CHECK-NEXT:    [[TMP1:%.*]] = fneg nsz <4 x double> [[X:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <4 x double> [[TMP1]], <4 x double> poison, <2 x i32> <i32 poison, i32 3>
+; CHECK-NEXT:    [[R:%.*]] = shufflevector <2 x double> [[Y:%.*]], <2 x double> [[TMP2]], <2 x i32> <i32 0, i32 3>
 ; CHECK-NEXT:    ret <2 x double> [[R]]
 ;
   %e = extractelement <4 x double> %x, i32 3
@@ -246,11 +320,17 @@ define <2 x double> @ext3_v4f64v2f64(<4 x double> %x, <2 x double> %y) {
 }
 
 define <4 x double> @ext1_v2f64v4f64_ins0(<2 x double> %x, <4 x double> %y) {
-; CHECK-LABEL: @ext1_v2f64v4f64_ins0(
-; CHECK-NEXT:    [[E:%.*]] = extractelement <2 x double> [[X:%.*]], i32 1
-; CHECK-NEXT:    [[N:%.*]] = fneg nsz double [[E]]
-; CHECK-NEXT:    [[R:%.*]] = insertelement <4 x double> [[Y:%.*]], double [[N]], i32 0
-; CHECK-NEXT:    ret <4 x double> [[R]]
+; SSE-LABEL: @ext1_v2f64v4f64_ins0(
+; SSE-NEXT:    [[TMP1:%.*]] = fneg nsz <2 x double> [[X:%.*]]
+; SSE-NEXT:    [[TMP2:%.*]] = shufflevector <2 x double> [[TMP1]], <2 x double> poison, <4 x i32> <i32 poison, i32 1, i32 poison, i32 poison>
+; SSE-NEXT:    [[R:%.*]] = shufflevector <4 x double> [[Y:%.*]], <4 x double> [[TMP2]], <4 x i32> <i32 5, i32 1, i32 2, i32 3>
+; SSE-NEXT:    ret <4 x double> [[R]]
+;
+; AVX-LABEL: @ext1_v2f64v4f64_ins0(
+; AVX-NEXT:    [[E:%.*]] = extractelement <2 x double> [[X:%.*]], i32 1
+; AVX-NEXT:    [[N:%.*]] = fneg nsz double [[E]]
+; AVX-NEXT:    [[R:%.*]] = insertelement <4 x double> [[Y:%.*]], double [[N]], i32 0
+; AVX-NEXT:    ret <4 x double> [[R]]
 ;
   %e = extractelement <2 x double> %x, i32 1
   %n = fneg nsz double %e


        


More information about the llvm-commits mailing list