llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT--> @llvm/pr-subscribers-vectorizers Author: Alexey Bataev (alexey-bataev) <details> <summary>Changes</summary> The bit tests compare and L, R, which sign-extends from the narrow type only if both L and R do, so its sign may change after narrowing. Fixes https://github.com/llvm/llvm-project/pull/233124#issuecomment-6382644843 Assisted-by: Cursor --- Full diff: https://github.com/llvm/llvm-project/pull/235287.diff 2 Files Affected: - (modified) llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp (+14-9) - (modified) llvm/test/Transforms/SLPVectorizer/X86/bit-test-narrowed-samesign.ll (+2-2) ``````````diff diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp index e23d074965678..496a511ba0f61 100644 --- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp +++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp @@ -18219,8 +18219,8 @@ BoUpSLP::getEntryCost(const TreeEntry *E, ArrayRef<Value *> VectorizedVals, cast<CmpInst>(E->getAltOp())->getPredicate(), CostKind, Op0Info, Op1Info, E->getAltOp()); } else { - Type *SrcTy = getWidenedType(E->getMainOp()->getOperand(0)->getType(), - VL.size()); + Type *SrcTy = + getWidenedType(E->getMainOp()->getOperand(0)->getType(), VL.size()); TTI::CastContextHint CCH = GetCastContextHint(E->getMainOp()->getOperand(0)); unsigned Opcode0 = E->getOpcode(); @@ -18262,12 +18262,12 @@ BoUpSLP::getEntryCost(const TreeEntry *E, ArrayRef<Value *> VectorizedVals, } // Same casts of the demoted operand, no need to shuffle. if (Opcode0 == Opcode1) - return VecCost + TTIRef.getCastInstrCost(Opcode0, VecTy, SrcTy, CCH, - CostKind); - VecCost += TTIRef.getCastInstrCost(Opcode0, VecTy, SrcTy, CCH, - CostKind, E->getMainOp()); - VecCost += TTIRef.getCastInstrCost(Opcode1, VecTy, SrcTy, CCH, - CostKind, E->getAltOp()); + return VecCost + + TTIRef.getCastInstrCost(Opcode0, VecTy, SrcTy, CCH, CostKind); + VecCost += TTIRef.getCastInstrCost(Opcode0, VecTy, SrcTy, CCH, CostKind, + E->getMainOp()); + VecCost += TTIRef.getCastInstrCost(Opcode1, VecTy, SrcTy, CCH, CostKind, + E->getAltOp()); } SmallVector<int> Mask; E->buildAltOpShuffleMask( @@ -24372,10 +24372,15 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) { // a known-zero narrow sign bit), or a converted lane's adjusted constant // flips the sign. if (auto *ICmp = dyn_cast<ICmpInst>(I); ICmp && It == MinBWs.end()) { + // The bit tests compare and L, R, which sign-extends from the narrow + // type only if both L and R do. + bool IsSigned0 = + GetOperandSignedness(0) && + (none_of(E->Scalars, IsaPred<TruncInst>) || GetOperandSignedness(1)); bool SignChange = ICmp->getOperand(0)->getType()->getScalarType() != VL0->getOperand(0)->getType()->getScalarType() && - ((!GetOperandSignedness(0) && + ((!IsSigned0 && !isKnownNonNegative(ICmp->getOperand(0), SimplifyQuery(*DL))) || (!GetOperandSignedness(1) && !isKnownNonNegative(ICmp->getOperand(1), SimplifyQuery(*DL)))); diff --git a/llvm/test/Transforms/SLPVectorizer/X86/bit-test-narrowed-samesign.ll b/llvm/test/Transforms/SLPVectorizer/X86/bit-test-narrowed-samesign.ll index 55beaeb6ee0c2..48783e7e96832 100644 --- a/llvm/test/Transforms/SLPVectorizer/X86/bit-test-narrowed-samesign.ll +++ b/llvm/test/Transforms/SLPVectorizer/X86/bit-test-narrowed-samesign.ll @@ -7,7 +7,7 @@ define void @samesign_ne_sext_mask_sign_bit(ptr noalias %pa, ptr noalias %dst) { ; CHECK-NEXT: [[ENTRY:.*:]] ; CHECK-NEXT: [[TMP0:%.*]] = load <4 x i8>, ptr [[PA]], align 1 ; CHECK-NEXT: [[TMP1:%.*]] = and <4 x i8> [[TMP0]], <i8 -128, i8 -128, i8 -128, i8 1> -; CHECK-NEXT: [[TMP2:%.*]] = icmp samesign ne <4 x i8> [[TMP1]], zeroinitializer +; CHECK-NEXT: [[TMP2:%.*]] = icmp ne <4 x i8> [[TMP1]], zeroinitializer ; CHECK-NEXT: [[TMP3:%.*]] = select <4 x i1> [[TMP2]], <4 x i32> splat (i32 3), <4 x i32> splat (i32 5) ; CHECK-NEXT: store <4 x i32> [[TMP3]], ptr [[DST]], align 4 ; CHECK-NEXT: ret void @@ -51,7 +51,7 @@ define void @samesign_eq_sext_mask_sign_bit(ptr noalias %pa, ptr noalias %dst) { ; CHECK-NEXT: [[ENTRY:.*:]] ; CHECK-NEXT: [[TMP0:%.*]] = load <4 x i8>, ptr [[PA]], align 1 ; CHECK-NEXT: [[TMP1:%.*]] = and <4 x i8> [[TMP0]], <i8 -128, i8 -128, i8 -128, i8 1> -; CHECK-NEXT: [[TMP2:%.*]] = icmp samesign eq <4 x i8> [[TMP1]], <i8 0, i8 0, i8 0, i8 1> +; CHECK-NEXT: [[TMP2:%.*]] = icmp eq <4 x i8> [[TMP1]], <i8 0, i8 0, i8 0, i8 1> ; CHECK-NEXT: [[TMP3:%.*]] = select <4 x i1> [[TMP2]], <4 x i32> splat (i32 3), <4 x i32> splat (i32 5) ; CHECK-NEXT: store <4 x i32> [[TMP3]], ptr [[DST]], align 4 ; CHECK-NEXT: ret void `````````` </details> https://github.com/llvm/llvm-project/pull/235287 _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
