Author: Krzysztof Parzyszek Date: 2020-12-12T13:27:22-06:00 New Revision: baf931a8427dc29532aa42460604f23bf9a39708
URL: https://github.com/llvm/llvm-project/commit/baf931a8427dc29532aa42460604f23bf9a39708 DIFF: https://github.com/llvm/llvm-project/commit/baf931a8427dc29532aa42460604f23bf9a39708.diff LOG: [Hexagon] Reconsider getMask fix, return original mask, convert later The getPayload/getMask/getPassThrough functions should return values that could be composed into a masked load/store without any additional type casts. The previous fix violated that. Instead, convert scalar mask to a vector right before rescaling. Added: Modified: llvm/lib/Target/Hexagon/HexagonVectorCombine.cpp Removed: ################################################################################ diff --git a/llvm/lib/Target/Hexagon/HexagonVectorCombine.cpp b/llvm/lib/Target/Hexagon/HexagonVectorCombine.cpp index fb716ef5bbc2..4c0c202be4be 100644 --- a/llvm/lib/Target/Hexagon/HexagonVectorCombine.cpp +++ b/llvm/lib/Target/Hexagon/HexagonVectorCombine.cpp @@ -416,8 +416,7 @@ auto AlignVectors::getMask(Value *Val) const -> Value * { int ElemCount = VecTy->getElementCount().getFixedValue(); return HVC.getFullValue(HVC.getBoolTy(ElemCount)); } - // For scalars, return a vector <1 x i1>. - return HVC.getFullValue(HVC.getBoolTy(1)); + return HVC.getFullValue(HVC.getBoolTy()); } auto AlignVectors::getPassThrough(Value *Val) const -> Value * { @@ -811,14 +810,24 @@ auto AlignVectors::realignGroup(const MoveGroup &Move) const -> bool { // Stores. ByteSpan ASpanV, ASpanM; + // Return a vector value corresponding to the input value Val: + // either <1 x Val> for scalar Val, or Val itself for vector Val. + auto MakeVec = [](IRBuilder<> &Builder, Value *Val) -> Value * { + Type *Ty = Val->getType(); + if (Ty->isVectorTy()) + return Val; + auto *VecTy = VectorType::get(Ty, 1, /*Scalable*/ false); + return Builder.CreateBitCast(Val, VecTy); + }; + for (int i = -1; i != NumSectors; ++i) { ByteSpan Section = VSpan.section(i * ScLen, ScLen).normalize(); Value *AccumV = UndefValue::get(SecTy); Value *AccumM = HVC.getNullValue(SecTy); for (ByteSpan::Block &S : Section) { Value *Pay = getPayload(S.Seg.Val); - Value *Mask = HVC.rescale(Builder, getMask(S.Seg.Val), Pay->getType(), - HVC.getByteTy()); + Value *Mask = HVC.rescale(Builder, MakeVec(Builder, getMask(S.Seg.Val)), + Pay->getType(), HVC.getByteTy()); AccumM = HVC.insertb(Builder, AccumM, HVC.vbytes(Builder, Mask), S.Seg.Start, S.Seg.Size, S.Pos); AccumV = HVC.insertb(Builder, AccumV, HVC.vbytes(Builder, Pay), _______________________________________________ llvm-branch-commits mailing list llvm-branch-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
