From 3720185c2869c146a5e889e05f4e62cee3dc314d Mon Sep 17 00:00:00 2001 From: Sanjay Patel Date: Sat, 12 Dec 2015 16:44:48 +0000 Subject: [PATCH] [InstCombine] canonicalize (bitcast (extractelement X)) --> (extractelement(bitcast X)) This change was discussed in D15392. It allows us to remove the fold that was added in: http://reviews.llvm.org/r255261 ...and it will allow us to generalize this fold: http://reviews.llvm.org/rL112232 while preserving the order of bitcast + extract that it produces and testing shows is better handled by the backend. Note that the existing check for "isVectorTy()" wasn't strong enough in general and specifically because: x86_mmx. It's not a vector, but it's not vectorizable either. So here we check VectorType::isValidElementType() directly before proceeding with the transform. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@255433 91177308-0d34-0410-b5e6-96231b3b80d8 --- .../InstCombine/InstCombineCasts.cpp | 45 +++++++------------ .../InstCombine/bitcast-vec-canon.ll | 4 +- 2 files changed, 19 insertions(+), 30 deletions(-) diff --git a/lib/Transforms/InstCombine/InstCombineCasts.cpp b/lib/Transforms/InstCombine/InstCombineCasts.cpp index dcd86db036b..b90e4d846bc 100644 --- a/lib/Transforms/InstCombine/InstCombineCasts.cpp +++ b/lib/Transforms/InstCombine/InstCombineCasts.cpp @@ -1715,40 +1715,29 @@ static Value *optimizeIntegerToVectorInsertions(BitCastInst &CI, return Result; } -/// Given a bitcasted source operand fed into an extract element instruction and -/// then bitcasted again to a scalar type, eliminate at least one bitcast by -/// changing the vector type of the extractelement instruction. -/// Example: -/// bitcast (extractelement (bitcast <2 x float> %X to <2 x i32>), 1) to float -/// ---> -/// extractelement <2 x float> %X, i32 1 -static Instruction *foldBitCastExtElt(BitCastInst &BitCast, InstCombiner &IC, - const DataLayout &DL) { - Type *DestType = BitCast.getType(); - if (DestType->isVectorTy()) - return nullptr; - +/// Canonicalize scalar bitcasts of extracted elements into a bitcast of the +/// vector followed by extract element. The backend tends to handle bitcasts of +/// vectors better than bitcasts of scalars because vector registers are +/// usually not type-specific like scalar integer or scalar floating-point. +static Instruction *canonicalizeBitCastExtElt(BitCastInst &BitCast, + InstCombiner &IC, + const DataLayout &DL) { // TODO: Create and use a pattern matcher for ExtractElementInst. auto *ExtElt = dyn_cast(BitCast.getOperand(0)); if (!ExtElt || !ExtElt->hasOneUse()) return nullptr; - Value *InnerBitCast = nullptr; - if (!match(ExtElt->getOperand(0), m_BitCast(m_Value(InnerBitCast)))) + // The bitcast must be to a vectorizable type, otherwise we can't make a new + // type to extract from. + Type *DestType = BitCast.getType(); + if (!VectorType::isValidElementType(DestType)) return nullptr; - // If the source is not a vector or its element type doesn't match the result - // type, bitcast it to a vector type that we can extract from. - Type *SourceType = InnerBitCast->getType(); - if (SourceType->getScalarType() != DestType) { - unsigned VecWidth = SourceType->getPrimitiveSizeInBits(); - unsigned DestWidth = DestType->getPrimitiveSizeInBits(); - unsigned NumElts = VecWidth / DestWidth; - SourceType = VectorType::get(DestType, NumElts); - InnerBitCast = IC.Builder->CreateBitCast(InnerBitCast, SourceType, "bc"); - } - - return ExtractElementInst::Create(InnerBitCast, ExtElt->getOperand(1)); + unsigned NumElts = ExtElt->getVectorOperandType()->getNumElements(); + auto *NewVecType = VectorType::get(DestType, NumElts); + auto *NewBC = IC.Builder->CreateBitCast(ExtElt->getVectorOperand(), + NewVecType, "bc"); + return ExtractElementInst::Create(NewBC, ExtElt->getIndexOperand()); } static Instruction *foldVecTruncToExtElt(Value *VecInput, Type *DestTy, @@ -1922,7 +1911,7 @@ Instruction *InstCombiner::visitBitCast(BitCastInst &CI) { } } - if (Instruction *I = foldBitCastExtElt(CI, *this, DL)) + if (Instruction *I = canonicalizeBitCastExtElt(CI, *this, DL)) return I; if (SrcTy->isPointerTy()) diff --git a/test/Transforms/InstCombine/bitcast-vec-canon.ll b/test/Transforms/InstCombine/bitcast-vec-canon.ll index e603036bfe8..97145221099 100644 --- a/test/Transforms/InstCombine/bitcast-vec-canon.ll +++ b/test/Transforms/InstCombine/bitcast-vec-canon.ll @@ -5,8 +5,8 @@ define double @a(<1 x i64> %y) { ret double %c ; CHECK-LABEL: @a( -; CHECK-NEXT: extractelement <1 x i64> %y, i32 0 -; CHECK-NEXT: bitcast i64 {{.*}} to double +; CHECK-NEXT: bitcast <1 x i64> %y to <1 x double> +; CHECK-NEXT: extractelement <1 x double> {{.*}}, i32 0 ; CHECK-NEXT: ret double } -- 2.34.1