shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
llvm-project/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
David Green 4b7913c357 [VectorCombine] Only consider shuffle uses with the same type. 
The backend getShuffleCosts do not currently handle shuffles that change
size very well. Limit the shuffles we collect to the same type to make
sure they do not cause issues as reported in D128732.
2022-07-16 13:23:39 +01:00

1674 lines
67 KiB
C++
Raw Blame History

//===------- VectorCombine.cpp - Optimize partial vector operations -------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This pass optimizes scalar/vector interactions using target cost models. The
// transforms implemented here may not fit in traditional loop-based or SLP
// vectorization passes.
//
//===----------------------------------------------------------------------===//
#include "llvm/Transforms/Vectorize/VectorCombine.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/BasicAliasAnalysis.h"
#include "llvm/Analysis/GlobalsModRef.h"
#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Analysis/VectorUtils.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Transforms/Vectorize.h"
#define DEBUG_TYPE "vector-combine"
#include "llvm/Transforms/Utils/InstructionWorklist.h"
using namespace llvm;
using namespace llvm::PatternMatch;
STATISTIC(NumVecLoad, "Number of vector loads formed");
STATISTIC(NumVecCmp, "Number of vector compares formed");
STATISTIC(NumVecBO, "Number of vector binops formed");
STATISTIC(NumVecCmpBO, "Number of vector compare + binop formed");
STATISTIC(NumShufOfBitcast, "Number of shuffles moved after bitcast");
STATISTIC(NumScalarBO, "Number of scalar binops formed");
STATISTIC(NumScalarCmp, "Number of scalar compares formed");
static cl::opt<bool> DisableVectorCombine(
"disable-vector-combine", cl::init(false), cl::Hidden,
cl::desc("Disable all vector combine transforms"));
static cl::opt<bool> DisableBinopExtractShuffle(
"disable-binop-extract-shuffle", cl::init(false), cl::Hidden,
cl::desc("Disable binop extract to shuffle transforms"));
static cl::opt<unsigned> MaxInstrsToScan(
"vector-combine-max-scan-instrs", cl::init(30), cl::Hidden,
cl::desc("Max number of instructions to scan for vector combining."));
static const unsigned InvalidIndex = std::numeric_limits<unsigned>::max();
namespace {
class VectorCombine {
public:
VectorCombine(Function &F, const TargetTransformInfo &TTI,
const DominatorTree &DT, AAResults &AA, AssumptionCache &AC,
bool ScalarizationOnly)
: F(F), Builder(F.getContext()), TTI(TTI), DT(DT), AA(AA), AC(AC),
ScalarizationOnly(ScalarizationOnly) {}
bool run();
private:
Function &F;
IRBuilder<> Builder;
const TargetTransformInfo &TTI;
const DominatorTree &DT;
AAResults &AA;
AssumptionCache &AC;
/// If true only perform scalarization combines and do not introduce new
/// vector operations.
bool ScalarizationOnly;
InstructionWorklist Worklist;
bool vectorizeLoadInsert(Instruction &I);
ExtractElementInst *getShuffleExtract(ExtractElementInst *Ext0,
ExtractElementInst *Ext1,
unsigned PreferredExtractIndex) const;
bool isExtractExtractCheap(ExtractElementInst *Ext0, ExtractElementInst *Ext1,
const Instruction &I,
ExtractElementInst *&ConvertToShuffle,
unsigned PreferredExtractIndex);
void foldExtExtCmp(ExtractElementInst *Ext0, ExtractElementInst *Ext1,
Instruction &I);
void foldExtExtBinop(ExtractElementInst *Ext0, ExtractElementInst *Ext1,
Instruction &I);
bool foldExtractExtract(Instruction &I);
bool foldBitcastShuf(Instruction &I);
bool scalarizeBinopOrCmp(Instruction &I);
bool foldExtractedCmps(Instruction &I);
bool foldSingleElementStore(Instruction &I);
bool scalarizeLoadExtract(Instruction &I);
bool foldShuffleOfBinops(Instruction &I);
bool foldShuffleFromReductions(Instruction &I);
bool foldSelectShuffle(Instruction &I, bool FromReduction = false);
void replaceValue(Value &Old, Value &New) {
Old.replaceAllUsesWith(&New);
if (auto *NewI = dyn_cast<Instruction>(&New)) {
New.takeName(&Old);
Worklist.pushUsersToWorkList(*NewI);
Worklist.pushValue(NewI);
}
Worklist.pushValue(&Old);
}
void eraseInstruction(Instruction &I) {
for (Value *Op : I.operands())
Worklist.pushValue(Op);
Worklist.remove(&I);
I.eraseFromParent();
}
};
} // namespace
bool VectorCombine::vectorizeLoadInsert(Instruction &I) {
// Match insert into fixed vector of scalar value.
// TODO: Handle non-zero insert index.
auto *Ty = dyn_cast<FixedVectorType>(I.getType());
Value *Scalar;
if (!Ty || !match(&I, m_InsertElt(m_Undef(), m_Value(Scalar), m_ZeroInt())) ||
!Scalar->hasOneUse())
return false;
// Optionally match an extract from another vector.
Value *X;
bool HasExtract = match(Scalar, m_ExtractElt(m_Value(X), m_ZeroInt()));
if (!HasExtract)
X = Scalar;
// Match source value as load of scalar or vector.
// Do not vectorize scalar load (widening) if atomic/volatile or under
// asan/hwasan/memtag/tsan. The widened load may load data from dirty regions
// or create data races non-existent in the source.
auto *Load = dyn_cast<LoadInst>(X);
if (!Load || !Load->isSimple() || !Load->hasOneUse() ||
Load->getFunction()->hasFnAttribute(Attribute::SanitizeMemTag) ||
mustSuppressSpeculation(*Load))
return false;
const DataLayout &DL = I.getModule()->getDataLayout();
Value *SrcPtr = Load->getPointerOperand()->stripPointerCasts();
assert(isa<PointerType>(SrcPtr->getType()) && "Expected a pointer type");
unsigned AS = Load->getPointerAddressSpace();
// We are potentially transforming byte-sized (8-bit) memory accesses, so make
// sure we have all of our type-based constraints in place for this target.
Type *ScalarTy = Scalar->getType();
uint64_t ScalarSize = ScalarTy->getPrimitiveSizeInBits();
unsigned MinVectorSize = TTI.getMinVectorRegisterBitWidth();
if (!ScalarSize || !MinVectorSize || MinVectorSize % ScalarSize != 0 ||
ScalarSize % 8 != 0)
return false;
// Check safety of replacing the scalar load with a larger vector load.
// We use minimal alignment (maximum flexibility) because we only care about
// the dereferenceable region. When calculating cost and creating a new op,
// we may use a larger value based on alignment attributes.
unsigned MinVecNumElts = MinVectorSize / ScalarSize;
auto *MinVecTy = VectorType::get(ScalarTy, MinVecNumElts, false);
unsigned OffsetEltIndex = 0;
Align Alignment = Load->getAlign();
if (!isSafeToLoadUnconditionally(SrcPtr, MinVecTy, Align(1), DL, Load, &DT)) {
// It is not safe to load directly from the pointer, but we can still peek
// through gep offsets and check if it safe to load from a base address with
// updated alignment. If it is, we can shuffle the element(s) into place
// after loading.
unsigned OffsetBitWidth = DL.getIndexTypeSizeInBits(SrcPtr->getType());
APInt Offset(OffsetBitWidth, 0);
SrcPtr = SrcPtr->stripAndAccumulateInBoundsConstantOffsets(DL, Offset);
// We want to shuffle the result down from a high element of a vector, so
// the offset must be positive.
if (Offset.isNegative())
return false;
// The offset must be a multiple of the scalar element to shuffle cleanly
// in the element's size.
uint64_t ScalarSizeInBytes = ScalarSize / 8;
if (Offset.urem(ScalarSizeInBytes) != 0)
return false;
// If we load MinVecNumElts, will our target element still be loaded?
OffsetEltIndex = Offset.udiv(ScalarSizeInBytes).getZExtValue();
if (OffsetEltIndex >= MinVecNumElts)
return false;
if (!isSafeToLoadUnconditionally(SrcPtr, MinVecTy, Align(1), DL, Load, &DT))
return false;
// Update alignment with offset value. Note that the offset could be negated
// to more accurately represent "(new) SrcPtr - Offset = (old) SrcPtr", but
// negation does not change the result of the alignment calculation.
Alignment = commonAlignment(Alignment, Offset.getZExtValue());
}
// Original pattern: insertelt undef, load [free casts of] PtrOp, 0
// Use the greater of the alignment on the load or its source pointer.
Alignment = std::max(SrcPtr->getPointerAlignment(DL), Alignment);
Type *LoadTy = Load->getType();
InstructionCost OldCost =
TTI.getMemoryOpCost(Instruction::Load, LoadTy, Alignment, AS);
APInt DemandedElts = APInt::getOneBitSet(MinVecNumElts, 0);
OldCost += TTI.getScalarizationOverhead(MinVecTy, DemandedElts,
/* Insert */ true, HasExtract);
// New pattern: load VecPtr
InstructionCost NewCost =
TTI.getMemoryOpCost(Instruction::Load, MinVecTy, Alignment, AS);
// Optionally, we are shuffling the loaded vector element(s) into place.
// For the mask set everything but element 0 to undef to prevent poison from
// propagating from the extra loaded memory. This will also optionally
// shrink/grow the vector from the loaded size to the output size.
// We assume this operation has no cost in codegen if there was no offset.
// Note that we could use freeze to avoid poison problems, but then we might
// still need a shuffle to change the vector size.
unsigned OutputNumElts = Ty->getNumElements();
SmallVector<int, 16> Mask(OutputNumElts, UndefMaskElem);
assert(OffsetEltIndex < MinVecNumElts && "Address offset too big");
Mask[0] = OffsetEltIndex;
if (OffsetEltIndex)
NewCost += TTI.getShuffleCost(TTI::SK_PermuteSingleSrc, MinVecTy, Mask);
// We can aggressively convert to the vector form because the backend can
// invert this transform if it does not result in a performance win.
if (OldCost < NewCost || !NewCost.isValid())
return false;
// It is safe and potentially profitable to load a vector directly:
// inselt undef, load Scalar, 0 --> load VecPtr
IRBuilder<> Builder(Load);
Value *CastedPtr = Builder.CreatePointerBitCastOrAddrSpaceCast(
SrcPtr, MinVecTy->getPointerTo(AS));
Value *VecLd = Builder.CreateAlignedLoad(MinVecTy, CastedPtr, Alignment);
VecLd = Builder.CreateShuffleVector(VecLd, Mask);
replaceValue(I, *VecLd);
++NumVecLoad;
return true;
}
/// Determine which, if any, of the inputs should be replaced by a shuffle
/// followed by extract from a different index.
ExtractElementInst *VectorCombine::getShuffleExtract(
ExtractElementInst *Ext0, ExtractElementInst *Ext1,
unsigned PreferredExtractIndex = InvalidIndex) const {
auto *Index0C = dyn_cast<ConstantInt>(Ext0->getIndexOperand());
auto *Index1C = dyn_cast<ConstantInt>(Ext1->getIndexOperand());
assert(Index0C && Index1C && "Expected constant extract indexes");
unsigned Index0 = Index0C->getZExtValue();
unsigned Index1 = Index1C->getZExtValue();
// If the extract indexes are identical, no shuffle is needed.
if (Index0 == Index1)
return nullptr;
Type *VecTy = Ext0->getVectorOperand()->getType();
assert(VecTy == Ext1->getVectorOperand()->getType() && "Need matching types");
InstructionCost Cost0 =
TTI.getVectorInstrCost(Ext0->getOpcode(), VecTy, Index0);
InstructionCost Cost1 =
TTI.getVectorInstrCost(Ext1->getOpcode(), VecTy, Index1);
// If both costs are invalid no shuffle is needed
if (!Cost0.isValid() && !Cost1.isValid())
return nullptr;
// We are extracting from 2 different indexes, so one operand must be shuffled
// before performing a vector operation and/or extract. The more expensive
// extract will be replaced by a shuffle.
if (Cost0 > Cost1)
return Ext0;
if (Cost1 > Cost0)
return Ext1;
// If the costs are equal and there is a preferred extract index, shuffle the
// opposite operand.
if (PreferredExtractIndex == Index0)
return Ext1;
if (PreferredExtractIndex == Index1)
return Ext0;
// Otherwise, replace the extract with the higher index.
return Index0 > Index1 ? Ext0 : Ext1;
}
/// Compare the relative costs of 2 extracts followed by scalar operation vs.
/// vector operation(s) followed by extract. Return true if the existing
/// instructions are cheaper than a vector alternative. Otherwise, return false
/// and if one of the extracts should be transformed to a shufflevector, set
/// \p ConvertToShuffle to that extract instruction.
bool VectorCombine::isExtractExtractCheap(ExtractElementInst *Ext0,
ExtractElementInst *Ext1,
const Instruction &I,
ExtractElementInst *&ConvertToShuffle,
unsigned PreferredExtractIndex) {
auto *Ext0IndexC = dyn_cast<ConstantInt>(Ext0->getOperand(1));
auto *Ext1IndexC = dyn_cast<ConstantInt>(Ext1->getOperand(1));
assert(Ext0IndexC && Ext1IndexC && "Expected constant extract indexes");
unsigned Opcode = I.getOpcode();
Type *ScalarTy = Ext0->getType();
auto *VecTy = cast<VectorType>(Ext0->getOperand(0)->getType());
InstructionCost ScalarOpCost, VectorOpCost;
// Get cost estimates for scalar and vector versions of the operation.
bool IsBinOp = Instruction::isBinaryOp(Opcode);
if (IsBinOp) {
ScalarOpCost = TTI.getArithmeticInstrCost(Opcode, ScalarTy);
VectorOpCost = TTI.getArithmeticInstrCost(Opcode, VecTy);
} else {
assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
"Expected a compare");
CmpInst::Predicate Pred = cast<CmpInst>(I).getPredicate();
ScalarOpCost = TTI.getCmpSelInstrCost(
Opcode, ScalarTy, CmpInst::makeCmpResultType(ScalarTy), Pred);
VectorOpCost = TTI.getCmpSelInstrCost(
Opcode, VecTy, CmpInst::makeCmpResultType(VecTy), Pred);
}
// Get cost estimates for the extract elements. These costs will factor into
// both sequences.
unsigned Ext0Index = Ext0IndexC->getZExtValue();
unsigned Ext1Index = Ext1IndexC->getZExtValue();
InstructionCost Extract0Cost =
TTI.getVectorInstrCost(Instruction::ExtractElement, VecTy, Ext0Index);
InstructionCost Extract1Cost =
TTI.getVectorInstrCost(Instruction::ExtractElement, VecTy, Ext1Index);
// A more expensive extract will always be replaced by a splat shuffle.
// For example, if Ext0 is more expensive:
// opcode (extelt V0, Ext0), (ext V1, Ext1) -->
// extelt (opcode (splat V0, Ext0), V1), Ext1
// TODO: Evaluate whether that always results in lowest cost. Alternatively,
// check the cost of creating a broadcast shuffle and shuffling both
// operands to element 0.
InstructionCost CheapExtractCost = std::min(Extract0Cost, Extract1Cost);
// Extra uses of the extracts mean that we include those costs in the
// vector total because those instructions will not be eliminated.
InstructionCost OldCost, NewCost;
if (Ext0->getOperand(0) == Ext1->getOperand(0) && Ext0Index == Ext1Index) {
// Handle a special case. If the 2 extracts are identical, adjust the
// formulas to account for that. The extra use charge allows for either the
// CSE'd pattern or an unoptimized form with identical values:
// opcode (extelt V, C), (extelt V, C) --> extelt (opcode V, V), C
bool HasUseTax = Ext0 == Ext1 ? !Ext0->hasNUses(2)
: !Ext0->hasOneUse() || !Ext1->hasOneUse();
OldCost = CheapExtractCost + ScalarOpCost;
NewCost = VectorOpCost + CheapExtractCost + HasUseTax * CheapExtractCost;
} else {
// Handle the general case. Each extract is actually a different value:
// opcode (extelt V0, C0), (extelt V1, C1) --> extelt (opcode V0, V1), C
OldCost = Extract0Cost + Extract1Cost + ScalarOpCost;
NewCost = VectorOpCost + CheapExtractCost +
!Ext0->hasOneUse() * Extract0Cost +
!Ext1->hasOneUse() * Extract1Cost;
}
ConvertToShuffle = getShuffleExtract(Ext0, Ext1, PreferredExtractIndex);
if (ConvertToShuffle) {
if (IsBinOp && DisableBinopExtractShuffle)
return true;
// If we are extracting from 2 different indexes, then one operand must be
// shuffled before performing the vector operation. The shuffle mask is
// undefined except for 1 lane that is being translated to the remaining
// extraction lane. Therefore, it is a splat shuffle. Ex:
// ShufMask = { undef, undef, 0, undef }
// TODO: The cost model has an option for a "broadcast" shuffle
// (splat-from-element-0), but no option for a more general splat.
NewCost +=
TTI.getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, VecTy);
}
// Aggressively form a vector op if the cost is equal because the transform
// may enable further optimization.
// Codegen can reverse this transform (scalarize) if it was not profitable.
return OldCost < NewCost;
}
/// Create a shuffle that translates (shifts) 1 element from the input vector
/// to a new element location.
static Value *createShiftShuffle(Value *Vec, unsigned OldIndex,
unsigned NewIndex, IRBuilder<> &Builder) {
// The shuffle mask is undefined except for 1 lane that is being translated
// to the new element index. Example for OldIndex == 2 and NewIndex == 0:
// ShufMask = { 2, undef, undef, undef }
auto *VecTy = cast<FixedVectorType>(Vec->getType());
SmallVector<int, 32> ShufMask(VecTy->getNumElements(), UndefMaskElem);
ShufMask[NewIndex] = OldIndex;
return Builder.CreateShuffleVector(Vec, ShufMask, "shift");
}
/// Given an extract element instruction with constant index operand, shuffle
/// the source vector (shift the scalar element) to a NewIndex for extraction.
/// Return null if the input can be constant folded, so that we are not creating
/// unnecessary instructions.
static ExtractElementInst *translateExtract(ExtractElementInst *ExtElt,
unsigned NewIndex,
IRBuilder<> &Builder) {
// Shufflevectors can only be created for fixed-width vectors.
if (!isa<FixedVectorType>(ExtElt->getOperand(0)->getType()))
return nullptr;
// If the extract can be constant-folded, this code is unsimplified. Defer
// to other passes to handle that.
Value *X = ExtElt->getVectorOperand();
Value *C = ExtElt->getIndexOperand();
assert(isa<ConstantInt>(C) && "Expected a constant index operand");
if (isa<Constant>(X))
return nullptr;
Value *Shuf = createShiftShuffle(X, cast<ConstantInt>(C)->getZExtValue(),
NewIndex, Builder);
return cast<ExtractElementInst>(Builder.CreateExtractElement(Shuf, NewIndex));
}
/// Try to reduce extract element costs by converting scalar compares to vector
/// compares followed by extract.
/// cmp (ext0 V0, C), (ext1 V1, C)
void VectorCombine::foldExtExtCmp(ExtractElementInst *Ext0,
ExtractElementInst *Ext1, Instruction &I) {
assert(isa<CmpInst>(&I) && "Expected a compare");
assert(cast<ConstantInt>(Ext0->getIndexOperand())->getZExtValue() ==
cast<ConstantInt>(Ext1->getIndexOperand())->getZExtValue() &&
"Expected matching constant extract indexes");
// cmp Pred (extelt V0, C), (extelt V1, C) --> extelt (cmp Pred V0, V1), C
++NumVecCmp;
CmpInst::Predicate Pred = cast<CmpInst>(&I)->getPredicate();
Value *V0 = Ext0->getVectorOperand(), *V1 = Ext1->getVectorOperand();
Value *VecCmp = Builder.CreateCmp(Pred, V0, V1);
Value *NewExt = Builder.CreateExtractElement(VecCmp, Ext0->getIndexOperand());
replaceValue(I, *NewExt);
}
/// Try to reduce extract element costs by converting scalar binops to vector
/// binops followed by extract.
/// bo (ext0 V0, C), (ext1 V1, C)
void VectorCombine::foldExtExtBinop(ExtractElementInst *Ext0,
ExtractElementInst *Ext1, Instruction &I) {
assert(isa<BinaryOperator>(&I) && "Expected a binary operator");
assert(cast<ConstantInt>(Ext0->getIndexOperand())->getZExtValue() ==
cast<ConstantInt>(Ext1->getIndexOperand())->getZExtValue() &&
"Expected matching constant extract indexes");
// bo (extelt V0, C), (extelt V1, C) --> extelt (bo V0, V1), C
++NumVecBO;
Value *V0 = Ext0->getVectorOperand(), *V1 = Ext1->getVectorOperand();
Value *VecBO =
Builder.CreateBinOp(cast<BinaryOperator>(&I)->getOpcode(), V0, V1);
// All IR flags are safe to back-propagate because any potential poison
// created in unused vector elements is discarded by the extract.
if (auto *VecBOInst = dyn_cast<Instruction>(VecBO))
VecBOInst->copyIRFlags(&I);
Value *NewExt = Builder.CreateExtractElement(VecBO, Ext0->getIndexOperand());
replaceValue(I, *NewExt);
}
/// Match an instruction with extracted vector operands.
bool VectorCombine::foldExtractExtract(Instruction &I) {
// It is not safe to transform things like div, urem, etc. because we may
// create undefined behavior when executing those on unknown vector elements.
if (!isSafeToSpeculativelyExecute(&I))
return false;
Instruction *I0, *I1;
CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
if (!match(&I, m_Cmp(Pred, m_Instruction(I0), m_Instruction(I1))) &&
!match(&I, m_BinOp(m_Instruction(I0), m_Instruction(I1))))
return false;
Value *V0, *V1;
uint64_t C0, C1;
if (!match(I0, m_ExtractElt(m_Value(V0), m_ConstantInt(C0))) ||
!match(I1, m_ExtractElt(m_Value(V1), m_ConstantInt(C1))) ||
V0->getType() != V1->getType())
return false;
// If the scalar value 'I' is going to be re-inserted into a vector, then try
// to create an extract to that same element. The extract/insert can be
// reduced to a "select shuffle".
// TODO: If we add a larger pattern match that starts from an insert, this
// probably becomes unnecessary.
auto *Ext0 = cast<ExtractElementInst>(I0);
auto *Ext1 = cast<ExtractElementInst>(I1);
uint64_t InsertIndex = InvalidIndex;
if (I.hasOneUse())
match(I.user_back(),
m_InsertElt(m_Value(), m_Value(), m_ConstantInt(InsertIndex)));
ExtractElementInst *ExtractToChange;
if (isExtractExtractCheap(Ext0, Ext1, I, ExtractToChange, InsertIndex))
return false;
if (ExtractToChange) {
unsigned CheapExtractIdx = ExtractToChange == Ext0 ? C1 : C0;
ExtractElementInst *NewExtract =
translateExtract(ExtractToChange, CheapExtractIdx, Builder);
if (!NewExtract)
return false;
if (ExtractToChange == Ext0)
Ext0 = NewExtract;
else
Ext1 = NewExtract;
}
if (Pred != CmpInst::BAD_ICMP_PREDICATE)
foldExtExtCmp(Ext0, Ext1, I);
else
foldExtExtBinop(Ext0, Ext1, I);
Worklist.push(Ext0);
Worklist.push(Ext1);
return true;
}
/// If this is a bitcast of a shuffle, try to bitcast the source vector to the
/// destination type followed by shuffle. This can enable further transforms by
/// moving bitcasts or shuffles together.
bool VectorCombine::foldBitcastShuf(Instruction &I) {
Value *V;
ArrayRef<int> Mask;
if (!match(&I, m_BitCast(
m_OneUse(m_Shuffle(m_Value(V), m_Undef(), m_Mask(Mask))))))
return false;
// 1) Do not fold bitcast shuffle for scalable type. First, shuffle cost for
// scalable type is unknown; Second, we cannot reason if the narrowed shuffle
// mask for scalable type is a splat or not.
// 2) Disallow non-vector casts and length-changing shuffles.
// TODO: We could allow any shuffle.
auto *DestTy = dyn_cast<FixedVectorType>(I.getType());
auto *SrcTy = dyn_cast<FixedVectorType>(V->getType());
if (!SrcTy || !DestTy || I.getOperand(0)->getType() != SrcTy)
return false;
unsigned DestNumElts = DestTy->getNumElements();
unsigned SrcNumElts = SrcTy->getNumElements();
SmallVector<int, 16> NewMask;
if (SrcNumElts <= DestNumElts) {
// The bitcast is from wide to narrow/equal elements. The shuffle mask can
// always be expanded to the equivalent form choosing narrower elements.
assert(DestNumElts % SrcNumElts == 0 && "Unexpected shuffle mask");
unsigned ScaleFactor = DestNumElts / SrcNumElts;
narrowShuffleMaskElts(ScaleFactor, Mask, NewMask);
} else {
// The bitcast is from narrow elements to wide elements. The shuffle mask
// must choose consecutive elements to allow casting first.
assert(SrcNumElts % DestNumElts == 0 && "Unexpected shuffle mask");
unsigned ScaleFactor = SrcNumElts / DestNumElts;
if (!widenShuffleMaskElts(ScaleFactor, Mask, NewMask))
return false;
}
// The new shuffle must not cost more than the old shuffle. The bitcast is
// moved ahead of the shuffle, so assume that it has the same cost as before.
InstructionCost DestCost = TTI.getShuffleCost(
TargetTransformInfo::SK_PermuteSingleSrc, DestTy, NewMask);
InstructionCost SrcCost =
TTI.getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, SrcTy, Mask);
if (DestCost > SrcCost || !DestCost.isValid())
return false;
// bitcast (shuf V, MaskC) --> shuf (bitcast V), MaskC'
++NumShufOfBitcast;
Value *CastV = Builder.CreateBitCast(V, DestTy);
Value *Shuf = Builder.CreateShuffleVector(CastV, NewMask);
replaceValue(I, *Shuf);
return true;
}
/// Match a vector binop or compare instruction with at least one inserted
/// scalar operand and convert to scalar binop/cmp followed by insertelement.
bool VectorCombine::scalarizeBinopOrCmp(Instruction &I) {
CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
Value *Ins0, *Ins1;
if (!match(&I, m_BinOp(m_Value(Ins0), m_Value(Ins1))) &&
!match(&I, m_Cmp(Pred, m_Value(Ins0), m_Value(Ins1))))
return false;
// Do not convert the vector condition of a vector select into a scalar
// condition. That may cause problems for codegen because of differences in
// boolean formats and register-file transfers.
// TODO: Can we account for that in the cost model?
bool IsCmp = Pred != CmpInst::Predicate::BAD_ICMP_PREDICATE;
if (IsCmp)
for (User *U : I.users())
if (match(U, m_Select(m_Specific(&I), m_Value(), m_Value())))
return false;
// Match against one or both scalar values being inserted into constant
// vectors:
// vec_op VecC0, (inselt VecC1, V1, Index)
// vec_op (inselt VecC0, V0, Index), VecC1
// vec_op (inselt VecC0, V0, Index), (inselt VecC1, V1, Index)
// TODO: Deal with mismatched index constants and variable indexes?
Constant *VecC0 = nullptr, *VecC1 = nullptr;
Value *V0 = nullptr, *V1 = nullptr;
uint64_t Index0 = 0, Index1 = 0;
if (!match(Ins0, m_InsertElt(m_Constant(VecC0), m_Value(V0),
m_ConstantInt(Index0))) &&
!match(Ins0, m_Constant(VecC0)))
return false;
if (!match(Ins1, m_InsertElt(m_Constant(VecC1), m_Value(V1),
m_ConstantInt(Index1))) &&
!match(Ins1, m_Constant(VecC1)))
return false;
bool IsConst0 = !V0;
bool IsConst1 = !V1;
if (IsConst0 && IsConst1)
return false;
if (!IsConst0 && !IsConst1 && Index0 != Index1)
return false;
// Bail for single insertion if it is a load.
// TODO: Handle this once getVectorInstrCost can cost for load/stores.
auto *I0 = dyn_cast_or_null<Instruction>(V0);
auto *I1 = dyn_cast_or_null<Instruction>(V1);
if ((IsConst0 && I1 && I1->mayReadFromMemory()) ||
(IsConst1 && I0 && I0->mayReadFromMemory()))
return false;
uint64_t Index = IsConst0 ? Index1 : Index0;
Type *ScalarTy = IsConst0 ? V1->getType() : V0->getType();
Type *VecTy = I.getType();
assert(VecTy->isVectorTy() &&
(IsConst0 || IsConst1 || V0->getType() == V1->getType()) &&
(ScalarTy->isIntegerTy() || ScalarTy->isFloatingPointTy() ||
ScalarTy->isPointerTy()) &&
"Unexpected types for insert element into binop or cmp");
unsigned Opcode = I.getOpcode();
InstructionCost ScalarOpCost, VectorOpCost;
if (IsCmp) {
CmpInst::Predicate Pred = cast<CmpInst>(I).getPredicate();
ScalarOpCost = TTI.getCmpSelInstrCost(
Opcode, ScalarTy, CmpInst::makeCmpResultType(ScalarTy), Pred);
VectorOpCost = TTI.getCmpSelInstrCost(
Opcode, VecTy, CmpInst::makeCmpResultType(VecTy), Pred);
} else {
ScalarOpCost = TTI.getArithmeticInstrCost(Opcode, ScalarTy);
VectorOpCost = TTI.getArithmeticInstrCost(Opcode, VecTy);
}
// Get cost estimate for the insert element. This cost will factor into
// both sequences.
InstructionCost InsertCost =
TTI.getVectorInstrCost(Instruction::InsertElement, VecTy, Index);
InstructionCost OldCost =
(IsConst0 ? 0 : InsertCost) + (IsConst1 ? 0 : InsertCost) + VectorOpCost;
InstructionCost NewCost = ScalarOpCost + InsertCost +
(IsConst0 ? 0 : !Ins0->hasOneUse() * InsertCost) +
(IsConst1 ? 0 : !Ins1->hasOneUse() * InsertCost);
// We want to scalarize unless the vector variant actually has lower cost.
if (OldCost < NewCost || !NewCost.isValid())
return false;
// vec_op (inselt VecC0, V0, Index), (inselt VecC1, V1, Index) -->
// inselt NewVecC, (scalar_op V0, V1), Index
if (IsCmp)
++NumScalarCmp;
else
++NumScalarBO;
// For constant cases, extract the scalar element, this should constant fold.
if (IsConst0)
V0 = ConstantExpr::getExtractElement(VecC0, Builder.getInt64(Index));
if (IsConst1)
V1 = ConstantExpr::getExtractElement(VecC1, Builder.getInt64(Index));
Value *Scalar =
IsCmp ? Builder.CreateCmp(Pred, V0, V1)
: Builder.CreateBinOp((Instruction::BinaryOps)Opcode, V0, V1);
Scalar->setName(I.getName() + ".scalar");
// All IR flags are safe to back-propagate. There is no potential for extra
// poison to be created by the scalar instruction.
if (auto *ScalarInst = dyn_cast<Instruction>(Scalar))
ScalarInst->copyIRFlags(&I);
// Fold the vector constants in the original vectors into a new base vector.
Value *NewVecC =
IsCmp ? Builder.CreateCmp(Pred, VecC0, VecC1)
: Builder.CreateBinOp((Instruction::BinaryOps)Opcode, VecC0, VecC1);
Value *Insert = Builder.CreateInsertElement(NewVecC, Scalar, Index);
replaceValue(I, *Insert);
return true;
}
/// Try to combine a scalar binop + 2 scalar compares of extracted elements of
/// a vector into vector operations followed by extract. Note: The SLP pass
/// may miss this pattern because of implementation problems.
bool VectorCombine::foldExtractedCmps(Instruction &I) {
// We are looking for a scalar binop of booleans.
// binop i1 (cmp Pred I0, C0), (cmp Pred I1, C1)
if (!I.isBinaryOp() || !I.getType()->isIntegerTy(1))
return false;
// The compare predicates should match, and each compare should have a
// constant operand.
// TODO: Relax the one-use constraints.
Value *B0 = I.getOperand(0), *B1 = I.getOperand(1);
Instruction *I0, *I1;
Constant *C0, *C1;
CmpInst::Predicate P0, P1;
if (!match(B0, m_OneUse(m_Cmp(P0, m_Instruction(I0), m_Constant(C0)))) ||
!match(B1, m_OneUse(m_Cmp(P1, m_Instruction(I1), m_Constant(C1)))) ||
P0 != P1)
return false;
// The compare operands must be extracts of the same vector with constant
// extract indexes.
// TODO: Relax the one-use constraints.
Value *X;
uint64_t Index0, Index1;
if (!match(I0, m_OneUse(m_ExtractElt(m_Value(X), m_ConstantInt(Index0)))) ||
!match(I1, m_OneUse(m_ExtractElt(m_Specific(X), m_ConstantInt(Index1)))))
return false;
auto *Ext0 = cast<ExtractElementInst>(I0);
auto *Ext1 = cast<ExtractElementInst>(I1);
ExtractElementInst *ConvertToShuf = getShuffleExtract(Ext0, Ext1);
if (!ConvertToShuf)
return false;
// The original scalar pattern is:
// binop i1 (cmp Pred (ext X, Index0), C0), (cmp Pred (ext X, Index1), C1)
CmpInst::Predicate Pred = P0;
unsigned CmpOpcode = CmpInst::isFPPredicate(Pred) ? Instruction::FCmp
: Instruction::ICmp;
auto *VecTy = dyn_cast<FixedVectorType>(X->getType());
if (!VecTy)
return false;
InstructionCost OldCost =
TTI.getVectorInstrCost(Ext0->getOpcode(), VecTy, Index0);
OldCost += TTI.getVectorInstrCost(Ext1->getOpcode(), VecTy, Index1);
OldCost +=
TTI.getCmpSelInstrCost(CmpOpcode, I0->getType(),
CmpInst::makeCmpResultType(I0->getType()), Pred) *
2;
OldCost += TTI.getArithmeticInstrCost(I.getOpcode(), I.getType());
// The proposed vector pattern is:
// vcmp = cmp Pred X, VecC
// ext (binop vNi1 vcmp, (shuffle vcmp, Index1)), Index0
int CheapIndex = ConvertToShuf == Ext0 ? Index1 : Index0;
int ExpensiveIndex = ConvertToShuf == Ext0 ? Index0 : Index1;
auto *CmpTy = cast<FixedVectorType>(CmpInst::makeCmpResultType(X->getType()));
InstructionCost NewCost = TTI.getCmpSelInstrCost(
CmpOpcode, X->getType(), CmpInst::makeCmpResultType(X->getType()), Pred);
SmallVector<int, 32> ShufMask(VecTy->getNumElements(), UndefMaskElem);
ShufMask[CheapIndex] = ExpensiveIndex;
NewCost += TTI.getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, CmpTy,
ShufMask);
NewCost += TTI.getArithmeticInstrCost(I.getOpcode(), CmpTy);
NewCost += TTI.getVectorInstrCost(Ext0->getOpcode(), CmpTy, CheapIndex);
// Aggressively form vector ops if the cost is equal because the transform
// may enable further optimization.
// Codegen can reverse this transform (scalarize) if it was not profitable.
if (OldCost < NewCost || !NewCost.isValid())
return false;
// Create a vector constant from the 2 scalar constants.
SmallVector<Constant *, 32> CmpC(VecTy->getNumElements(),
UndefValue::get(VecTy->getElementType()));
CmpC[Index0] = C0;
CmpC[Index1] = C1;
Value *VCmp = Builder.CreateCmp(Pred, X, ConstantVector::get(CmpC));
Value *Shuf = createShiftShuffle(VCmp, ExpensiveIndex, CheapIndex, Builder);
Value *VecLogic = Builder.CreateBinOp(cast<BinaryOperator>(I).getOpcode(),
VCmp, Shuf);
Value *NewExt = Builder.CreateExtractElement(VecLogic, CheapIndex);
replaceValue(I, *NewExt);
++NumVecCmpBO;
return true;
}
// Check if memory loc modified between two instrs in the same BB
static bool isMemModifiedBetween(BasicBlock::iterator Begin,
BasicBlock::iterator End,
const MemoryLocation &Loc, AAResults &AA) {
unsigned NumScanned = 0;
return std::any_of(Begin, End, [&](const Instruction &Instr) {
return isModSet(AA.getModRefInfo(&Instr, Loc)) ||
++NumScanned > MaxInstrsToScan;
});
}
/// Helper class to indicate whether a vector index can be safely scalarized and
/// if a freeze needs to be inserted.
class ScalarizationResult {
enum class StatusTy { Unsafe, Safe, SafeWithFreeze };
StatusTy Status;
Value *ToFreeze;
ScalarizationResult(StatusTy Status, Value *ToFreeze = nullptr)
: Status(Status), ToFreeze(ToFreeze) {}
public:
ScalarizationResult(const ScalarizationResult &Other) = default;
~ScalarizationResult() {
assert(!ToFreeze && "freeze() not called with ToFreeze being set");
}
static ScalarizationResult unsafe() { return {StatusTy::Unsafe}; }
static ScalarizationResult safe() { return {StatusTy::Safe}; }
static ScalarizationResult safeWithFreeze(Value *ToFreeze) {
return {StatusTy::SafeWithFreeze, ToFreeze};
}
/// Returns true if the index can be scalarize without requiring a freeze.
bool isSafe() const { return Status == StatusTy::Safe; }
/// Returns true if the index cannot be scalarized.
bool isUnsafe() const { return Status == StatusTy::Unsafe; }
/// Returns true if the index can be scalarize, but requires inserting a
/// freeze.
bool isSafeWithFreeze() const { return Status == StatusTy::SafeWithFreeze; }
/// Reset the state of Unsafe and clear ToFreze if set.
void discard() {
ToFreeze = nullptr;
Status = StatusTy::Unsafe;
}
/// Freeze the ToFreeze and update the use in \p User to use it.
void freeze(IRBuilder<> &Builder, Instruction &UserI) {
assert(isSafeWithFreeze() &&
"should only be used when freezing is required");
assert(is_contained(ToFreeze->users(), &UserI) &&
"UserI must be a user of ToFreeze");
IRBuilder<>::InsertPointGuard Guard(Builder);
Builder.SetInsertPoint(cast<Instruction>(&UserI));
Value *Frozen =
Builder.CreateFreeze(ToFreeze, ToFreeze->getName() + ".frozen");
for (Use &U : make_early_inc_range((UserI.operands())))
if (U.get() == ToFreeze)
U.set(Frozen);
ToFreeze = nullptr;
}
};
/// Check if it is legal to scalarize a memory access to \p VecTy at index \p
/// Idx. \p Idx must access a valid vector element.
static ScalarizationResult canScalarizeAccess(FixedVectorType *VecTy,
Value *Idx, Instruction *CtxI,
AssumptionCache &AC,
const DominatorTree &DT) {
if (auto *C = dyn_cast<ConstantInt>(Idx)) {
if (C->getValue().ult(VecTy->getNumElements()))
return ScalarizationResult::safe();
return ScalarizationResult::unsafe();
}
unsigned IntWidth = Idx->getType()->getScalarSizeInBits();
APInt Zero(IntWidth, 0);
APInt MaxElts(IntWidth, VecTy->getNumElements());
ConstantRange ValidIndices(Zero, MaxElts);
ConstantRange IdxRange(IntWidth, true);
if (isGuaranteedNotToBePoison(Idx, &AC)) {
if (ValidIndices.contains(computeConstantRange(Idx, /* ForSigned */ false,
true, &AC, CtxI, &DT)))
return ScalarizationResult::safe();
return ScalarizationResult::unsafe();
}
// If the index may be poison, check if we can insert a freeze before the
// range of the index is restricted.
Value *IdxBase;
ConstantInt *CI;
if (match(Idx, m_And(m_Value(IdxBase), m_ConstantInt(CI)))) {
IdxRange = IdxRange.binaryAnd(CI->getValue());
} else if (match(Idx, m_URem(m_Value(IdxBase), m_ConstantInt(CI)))) {
IdxRange = IdxRange.urem(CI->getValue());
}
if (ValidIndices.contains(IdxRange))
return ScalarizationResult::safeWithFreeze(IdxBase);
return ScalarizationResult::unsafe();
}
/// The memory operation on a vector of \p ScalarType had alignment of
/// \p VectorAlignment. Compute the maximal, but conservatively correct,
/// alignment that will be valid for the memory operation on a single scalar
/// element of the same type with index \p Idx.
static Align computeAlignmentAfterScalarization(Align VectorAlignment,
Type *ScalarType, Value *Idx,
const DataLayout &DL) {
if (auto *C = dyn_cast<ConstantInt>(Idx))
return commonAlignment(VectorAlignment,
C->getZExtValue() * DL.getTypeStoreSize(ScalarType));
return commonAlignment(VectorAlignment, DL.getTypeStoreSize(ScalarType));
}
// Combine patterns like:
// %0 = load <4 x i32>, <4 x i32>* %a
// %1 = insertelement <4 x i32> %0, i32 %b, i32 1
// store <4 x i32> %1, <4 x i32>* %a
// to:
// %0 = bitcast <4 x i32>* %a to i32*
// %1 = getelementptr inbounds i32, i32* %0, i64 0, i64 1
// store i32 %b, i32* %1
bool VectorCombine::foldSingleElementStore(Instruction &I) {
StoreInst *SI = dyn_cast<StoreInst>(&I);
if (!SI || !SI->isSimple() ||
!isa<FixedVectorType>(SI->getValueOperand()->getType()))
return false;
// TODO: Combine more complicated patterns (multiple insert) by referencing
// TargetTransformInfo.
Instruction *Source;
Value *NewElement;
Value *Idx;
if (!match(SI->getValueOperand(),
m_InsertElt(m_Instruction(Source), m_Value(NewElement),
m_Value(Idx))))
return false;
if (auto *Load = dyn_cast<LoadInst>(Source)) {
auto VecTy = cast<FixedVectorType>(SI->getValueOperand()->getType());
const DataLayout &DL = I.getModule()->getDataLayout();
Value *SrcAddr = Load->getPointerOperand()->stripPointerCasts();
// Don't optimize for atomic/volatile load or store. Ensure memory is not
// modified between, vector type matches store size, and index is inbounds.
if (!Load->isSimple() || Load->getParent() != SI->getParent() ||
!DL.typeSizeEqualsStoreSize(Load->getType()) ||
SrcAddr != SI->getPointerOperand()->stripPointerCasts())
return false;
auto ScalarizableIdx = canScalarizeAccess(VecTy, Idx, Load, AC, DT);
if (ScalarizableIdx.isUnsafe() ||
isMemModifiedBetween(Load->getIterator(), SI->getIterator(),
MemoryLocation::get(SI), AA))
return false;
if (ScalarizableIdx.isSafeWithFreeze())
ScalarizableIdx.freeze(Builder, *cast<Instruction>(Idx));
Value *GEP = Builder.CreateInBoundsGEP(
SI->getValueOperand()->getType(), SI->getPointerOperand(),
{ConstantInt::get(Idx->getType(), 0), Idx});
StoreInst *NSI = Builder.CreateStore(NewElement, GEP);
NSI->copyMetadata(*SI);
Align ScalarOpAlignment = computeAlignmentAfterScalarization(
std::max(SI->getAlign(), Load->getAlign()), NewElement->getType(), Idx,
DL);
NSI->setAlignment(ScalarOpAlignment);
replaceValue(I, *NSI);
eraseInstruction(I);
return true;
}
return false;
}
/// Try to scalarize vector loads feeding extractelement instructions.
bool VectorCombine::scalarizeLoadExtract(Instruction &I) {
Value *Ptr;
if (!match(&I, m_Load(m_Value(Ptr))))
return false;
auto *LI = cast<LoadInst>(&I);
const DataLayout &DL = I.getModule()->getDataLayout();
if (LI->isVolatile() || !DL.typeSizeEqualsStoreSize(LI->getType()))
return false;
auto *FixedVT = dyn_cast<FixedVectorType>(LI->getType());
if (!FixedVT)
return false;
InstructionCost OriginalCost =
TTI.getMemoryOpCost(Instruction::Load, LI->getType(), LI->getAlign(),
LI->getPointerAddressSpace());
InstructionCost ScalarizedCost = 0;
Instruction *LastCheckedInst = LI;
unsigned NumInstChecked = 0;
// Check if all users of the load are extracts with no memory modifications
// between the load and the extract. Compute the cost of both the original
// code and the scalarized version.
for (User *U : LI->users()) {
auto *UI = dyn_cast<ExtractElementInst>(U);
if (!UI || UI->getParent() != LI->getParent())
return false;
if (!isGuaranteedNotToBePoison(UI->getOperand(1), &AC, LI, &DT))
return false;
// Check if any instruction between the load and the extract may modify
// memory.
if (LastCheckedInst->comesBefore(UI)) {
for (Instruction &I :
make_range(std::next(LI->getIterator()), UI->getIterator())) {
// Bail out if we reached the check limit or the instruction may write
// to memory.
if (NumInstChecked == MaxInstrsToScan || I.mayWriteToMemory())
return false;
NumInstChecked++;
}
LastCheckedInst = UI;
}
auto ScalarIdx = canScalarizeAccess(FixedVT, UI->getOperand(1), &I, AC, DT);
if (!ScalarIdx.isSafe()) {
// TODO: Freeze index if it is safe to do so.
ScalarIdx.discard();
return false;
}
auto *Index = dyn_cast<ConstantInt>(UI->getOperand(1));
OriginalCost +=
TTI.getVectorInstrCost(Instruction::ExtractElement, LI->getType(),
Index ? Index->getZExtValue() : -1);
ScalarizedCost +=
TTI.getMemoryOpCost(Instruction::Load, FixedVT->getElementType(),
Align(1), LI->getPointerAddressSpace());
ScalarizedCost += TTI.getAddressComputationCost(FixedVT->getElementType());
}
if (ScalarizedCost >= OriginalCost)
return false;
// Replace extracts with narrow scalar loads.
for (User *U : LI->users()) {
auto *EI = cast<ExtractElementInst>(U);
Builder.SetInsertPoint(EI);
Value *Idx = EI->getOperand(1);
Value *GEP =
Builder.CreateInBoundsGEP(FixedVT, Ptr, {Builder.getInt32(0), Idx});
auto *NewLoad = cast<LoadInst>(Builder.CreateLoad(
FixedVT->getElementType(), GEP, EI->getName() + ".scalar"));
Align ScalarOpAlignment = computeAlignmentAfterScalarization(
LI->getAlign(), FixedVT->getElementType(), Idx, DL);
NewLoad->setAlignment(ScalarOpAlignment);
replaceValue(*EI, *NewLoad);
}
return true;
}
/// Try to convert "shuffle (binop), (binop)" with a shared binop operand into
/// "binop (shuffle), (shuffle)".
bool VectorCombine::foldShuffleOfBinops(Instruction &I) {
auto *VecTy = dyn_cast<FixedVectorType>(I.getType());
if (!VecTy)
return false;
BinaryOperator *B0, *B1;
ArrayRef<int> Mask;
if (!match(&I, m_Shuffle(m_OneUse(m_BinOp(B0)), m_OneUse(m_BinOp(B1)),
m_Mask(Mask))) ||
B0->getOpcode() != B1->getOpcode() || B0->getType() != VecTy)
return false;
// Try to replace a binop with a shuffle if the shuffle is not costly.
// The new shuffle will choose from a single, common operand, so it may be
// cheaper than the existing two-operand shuffle.
SmallVector<int> UnaryMask = createUnaryMask(Mask, Mask.size());
Instruction::BinaryOps Opcode = B0->getOpcode();
InstructionCost BinopCost = TTI.getArithmeticInstrCost(Opcode, VecTy);
InstructionCost ShufCost = TTI.getShuffleCost(
TargetTransformInfo::SK_PermuteSingleSrc, VecTy, UnaryMask);
if (ShufCost > BinopCost)
return false;
// If we have something like "add X, Y" and "add Z, X", swap ops to match.
Value *X = B0->getOperand(0), *Y = B0->getOperand(1);
Value *Z = B1->getOperand(0), *W = B1->getOperand(1);
if (BinaryOperator::isCommutative(Opcode) && X != Z && Y != W)
std::swap(X, Y);
Value *Shuf0, *Shuf1;
if (X == Z) {
// shuf (bo X, Y), (bo X, W) --> bo (shuf X), (shuf Y, W)
Shuf0 = Builder.CreateShuffleVector(X, UnaryMask);
Shuf1 = Builder.CreateShuffleVector(Y, W, Mask);
} else if (Y == W) {
// shuf (bo X, Y), (bo Z, Y) --> bo (shuf X, Z), (shuf Y)
Shuf0 = Builder.CreateShuffleVector(X, Z, Mask);
Shuf1 = Builder.CreateShuffleVector(Y, UnaryMask);
} else {
return false;
}
Value *NewBO = Builder.CreateBinOp(Opcode, Shuf0, Shuf1);
// Intersect flags from the old binops.
if (auto *NewInst = dyn_cast<Instruction>(NewBO)) {
NewInst->copyIRFlags(B0);
NewInst->andIRFlags(B1);
}
replaceValue(I, *NewBO);
return true;
}
/// Given a commutative reduction, the order of the input lanes does not alter
/// the results. We can use this to remove certain shuffles feeding the
/// reduction, removing the need to shuffle at all.
bool VectorCombine::foldShuffleFromReductions(Instruction &I) {
auto *II = dyn_cast<IntrinsicInst>(&I);
if (!II)
return false;
switch (II->getIntrinsicID()) {
case Intrinsic::vector_reduce_add:
case Intrinsic::vector_reduce_mul:
case Intrinsic::vector_reduce_and:
case Intrinsic::vector_reduce_or:
case Intrinsic::vector_reduce_xor:
case Intrinsic::vector_reduce_smin:
case Intrinsic::vector_reduce_smax:
case Intrinsic::vector_reduce_umin:
case Intrinsic::vector_reduce_umax:
break;
default:
return false;
}
// Find all the inputs when looking through operations that do not alter the
// lane order (binops, for example). Currently we look for a single shuffle,
// and can ignore splat values.
std::queue<Value *> Worklist;
SmallPtrSet<Value *, 4> Visited;
ShuffleVectorInst *Shuffle = nullptr;
if (auto *Op = dyn_cast<Instruction>(I.getOperand(0)))
Worklist.push(Op);
while (!Worklist.empty()) {
Value *CV = Worklist.front();
Worklist.pop();
if (Visited.contains(CV))
continue;
// Splats don't change the order, so can be safely ignored.
if (isSplatValue(CV))
continue;
Visited.insert(CV);
if (auto *CI = dyn_cast<Instruction>(CV)) {
if (CI->isBinaryOp()) {
for (auto *Op : CI->operand_values())
Worklist.push(Op);
continue;
} else if (auto *SV = dyn_cast<ShuffleVectorInst>(CI)) {
if (Shuffle && Shuffle != SV)
return false;
Shuffle = SV;
continue;
}
}
// Anything else is currently an unknown node.
return false;
}
if (!Shuffle)
return false;
// Check all uses of the binary ops and shuffles are also included in the
// lane-invariant operations (Visited should be the list of lanewise
// instructions, including the shuffle that we found).
for (auto *V : Visited)
for (auto *U : V->users())
if (!Visited.contains(U) && U != &I)
return false;
FixedVectorType *VecType =
dyn_cast<FixedVectorType>(II->getOperand(0)->getType());
if (!VecType)
return false;
FixedVectorType *ShuffleInputType =
dyn_cast<FixedVectorType>(Shuffle->getOperand(0)->getType());
if (!ShuffleInputType)
return false;
int NumInputElts = ShuffleInputType->getNumElements();
// Find the mask from sorting the lanes into order. This is most likely to
// become a identity or concat mask. Undef elements are pushed to the end.
SmallVector<int> ConcatMask;
Shuffle->getShuffleMask(ConcatMask);
sort(ConcatMask, [](int X, int Y) { return (unsigned)X < (unsigned)Y; });
bool UsesSecondVec =
any_of(ConcatMask, [&](int M) { return M >= NumInputElts; });
InstructionCost OldCost = TTI.getShuffleCost(
UsesSecondVec ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc, VecType,
Shuffle->getShuffleMask());
InstructionCost NewCost = TTI.getShuffleCost(
UsesSecondVec ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc, VecType,
ConcatMask);
LLVM_DEBUG(dbgs() << "Found a reduction feeding from a shuffle: " << *Shuffle
<< "\n");
LLVM_DEBUG(dbgs() << " OldCost: " << OldCost << " vs NewCost: " << NewCost
<< "\n");
if (NewCost < OldCost) {
Builder.SetInsertPoint(Shuffle);
Value *NewShuffle = Builder.CreateShuffleVector(
Shuffle->getOperand(0), Shuffle->getOperand(1), ConcatMask);
LLVM_DEBUG(dbgs() << "Created new shuffle: " << *NewShuffle << "\n");
replaceValue(*Shuffle, *NewShuffle);
}
// See if we can re-use foldSelectShuffle, getting it to reduce the size of
// the shuffle into a nicer order, as it can ignore the order of the shuffles.
return foldSelectShuffle(*Shuffle, true);
}
/// This method looks for groups of shuffles acting on binops, of the form:
/// %x = shuffle ...
/// %y = shuffle ...
/// %a = binop %x, %y
/// %b = binop %x, %y
/// shuffle %a, %b, selectmask
/// We may, especially if the shuffle is wider than legal, be able to convert
/// the shuffle to a form where only parts of a and b need to be computed. On
/// architectures with no obvious "select" shuffle, this can reduce the total
/// number of operations if the target reports them as cheaper.
bool VectorCombine::foldSelectShuffle(Instruction &I, bool FromReduction) {
auto *SVI = dyn_cast<ShuffleVectorInst>(&I);
auto *VT = dyn_cast<FixedVectorType>(I.getType());
if (!SVI || !VT)
return false;
auto *Op0 = dyn_cast<Instruction>(SVI->getOperand(0));
auto *Op1 = dyn_cast<Instruction>(SVI->getOperand(1));
if (!Op0 || !Op1 || Op0 == Op1 || !Op0->isBinaryOp() || !Op1->isBinaryOp() ||
VT != Op0->getType())
return false;
auto *SVI0A = dyn_cast<Instruction>(Op0->getOperand(0));
auto *SVI0B = dyn_cast<Instruction>(Op0->getOperand(1));
auto *SVI1A = dyn_cast<Instruction>(Op1->getOperand(0));
auto *SVI1B = dyn_cast<Instruction>(Op1->getOperand(1));
SmallPtrSet<Instruction *, 4> InputShuffles({SVI0A, SVI0B, SVI1A, SVI1B});
auto checkSVNonOpUses = [&](Instruction *I) {
if (!I || I->getOperand(0)->getType() != VT)
return true;
return any_of(I->users(), [&](User *U) {
return U != Op0 && U != Op1 &&
!(isa<ShuffleVectorInst>(U) &&
(InputShuffles.contains(cast<Instruction>(U)) ||
isInstructionTriviallyDead(cast<Instruction>(U))));
});
};
if (checkSVNonOpUses(SVI0A) || checkSVNonOpUses(SVI0B) ||
checkSVNonOpUses(SVI1A) || checkSVNonOpUses(SVI1B))
return false;
// Collect all the uses that are shuffles that we can transform together. We
// may not have a single shuffle, but a group that can all be transformed
// together profitably.
SmallVector<ShuffleVectorInst *> Shuffles;
auto collectShuffles = [&](Instruction *I) {
for (auto *U : I->users()) {
auto *SV = dyn_cast<ShuffleVectorInst>(U);
if (!SV || SV->getType() != VT)
return false;
if ((SV->getOperand(0) != Op0 && SV->getOperand(0) != Op1) ||
(SV->getOperand(1) != Op0 && SV->getOperand(1) != Op1))
return false;
if (!llvm::is_contained(Shuffles, SV))
Shuffles.push_back(SV);
}
return true;
};
if (!collectShuffles(Op0) || !collectShuffles(Op1))
return false;
// From a reduction, we need to be processing a single shuffle, otherwise the
// other uses will not be lane-invariant.
if (FromReduction && Shuffles.size() > 1)
return false;
// Add any shuffle uses for the shuffles we have found, to include them in our
// cost calculations.
if (!FromReduction) {
for (ShuffleVectorInst *SV : Shuffles) {
for (auto U : SV->users()) {
ShuffleVectorInst *SSV = dyn_cast<ShuffleVectorInst>(U);
if (SSV && isa<UndefValue>(SSV->getOperand(1)) && SSV->getType() == VT)
Shuffles.push_back(SSV);
}
}
}
// For each of the output shuffles, we try to sort all the first vector
// elements to the beginning, followed by the second array elements at the
// end. If the binops are legalized to smaller vectors, this may reduce total
// number of binops. We compute the ReconstructMask mask needed to convert
// back to the original lane order.
SmallVector<std::pair<int, int>> V1, V2;
SmallVector<SmallVector<int>> OrigReconstructMasks;
int MaxV1Elt = 0, MaxV2Elt = 0;
unsigned NumElts = VT->getNumElements();
for (ShuffleVectorInst *SVN : Shuffles) {
SmallVector<int> Mask;
SVN->getShuffleMask(Mask);
// Check the operands are the same as the original, or reversed (in which
// case we need to commute the mask).
Value *SVOp0 = SVN->getOperand(0);
Value *SVOp1 = SVN->getOperand(1);
if (isa<UndefValue>(SVOp1)) {
auto *SSV = cast<ShuffleVectorInst>(SVOp0);
SVOp0 = SSV->getOperand(0);
SVOp1 = SSV->getOperand(1);
for (unsigned I = 0, E = Mask.size(); I != E; I++) {
if (Mask[I] >= static_cast<int>(SSV->getShuffleMask().size()))
return false;
Mask[I] = Mask[I] < 0 ? Mask[I] : SSV->getMaskValue(Mask[I]);
}
}
if (SVOp0 == Op1 && SVOp1 == Op0) {
std::swap(SVOp0, SVOp1);
ShuffleVectorInst::commuteShuffleMask(Mask, NumElts);
}
if (SVOp0 != Op0 || SVOp1 != Op1)
return false;
// Calculate the reconstruction mask for this shuffle, as the mask needed to
// take the packed values from Op0/Op1 and reconstructing to the original
// order.
SmallVector<int> ReconstructMask;
for (unsigned I = 0; I < Mask.size(); I++) {
if (Mask[I] < 0) {
ReconstructMask.push_back(-1);
} else if (Mask[I] < static_cast<int>(NumElts)) {
MaxV1Elt = std::max(MaxV1Elt, Mask[I]);
auto It = find_if(V1, [&](const std::pair<int, int> &A) {
return Mask[I] == A.first;
});
if (It != V1.end())
ReconstructMask.push_back(It - V1.begin());
else {
ReconstructMask.push_back(V1.size());
V1.emplace_back(Mask[I], V1.size());
}
} else {
MaxV2Elt = std::max<int>(MaxV2Elt, Mask[I] - NumElts);
auto It = find_if(V2, [&](const std::pair<int, int> &A) {
return Mask[I] - static_cast<int>(NumElts) == A.first;
});
if (It != V2.end())
ReconstructMask.push_back(NumElts + It - V2.begin());
else {
ReconstructMask.push_back(NumElts + V2.size());
V2.emplace_back(Mask[I] - NumElts, NumElts + V2.size());
}
}
}
// For reductions, we know that the lane ordering out doesn't alter the
// result. In-order can help simplify the shuffle away.
if (FromReduction)
sort(ReconstructMask);
OrigReconstructMasks.push_back(std::move(ReconstructMask));
}
// If the Maximum element used from V1 and V2 are not larger than the new
// vectors, the vectors are already packes and performing the optimization
// again will likely not help any further. This also prevents us from getting
// stuck in a cycle in case the costs do not also rule it out.
if (V1.empty() || V2.empty() ||
(MaxV1Elt == static_cast<int>(V1.size()) - 1 &&
MaxV2Elt == static_cast<int>(V2.size()) - 1))
return false;
// GetBaseMaskValue takes one of the inputs, which may either be a shuffle, a
// shuffle of another shuffle, or not a shuffle (that is treated like a
// identity shuffle).
auto GetBaseMaskValue = [&](Instruction *I, int M) {
auto *SV = dyn_cast<ShuffleVectorInst>(I);
if (!SV)
return M;
if (isa<UndefValue>(SV->getOperand(1)))
if (auto *SSV = dyn_cast<ShuffleVectorInst>(SV->getOperand(0)))
if (InputShuffles.contains(SSV))
return SSV->getMaskValue(SV->getMaskValue(M));
return SV->getMaskValue(M);
};
// Attempt to sort the inputs my ascending mask values to make simpler input
// shuffles and push complex shuffles down to the uses. We sort on the first
// of the two input shuffle orders, to try and get at least one input into a
// nice order.
auto SortBase = [&](Instruction *A, std::pair<int, int> X,
std::pair<int, int> Y) {
int MXA = GetBaseMaskValue(A, X.first);
int MYA = GetBaseMaskValue(A, Y.first);
return MXA < MYA;
};
stable_sort(V1, [&](std::pair<int, int> A, std::pair<int, int> B) {
return SortBase(SVI0A, A, B);
});
stable_sort(V2, [&](std::pair<int, int> A, std::pair<int, int> B) {
return SortBase(SVI1A, A, B);
});
// Calculate our ReconstructMasks from the OrigReconstructMasks and the
// modified order of the input shuffles.
SmallVector<SmallVector<int>> ReconstructMasks;
for (auto Mask : OrigReconstructMasks) {
SmallVector<int> ReconstructMask;
for (int M : Mask) {
auto FindIndex = [](const SmallVector<std::pair<int, int>> &V, int M) {
auto It = find_if(V, [M](auto A) { return A.second == M; });
assert(It != V.end() && "Expected all entries in Mask");
return std::distance(V.begin(), It);
};
if (M < 0)
ReconstructMask.push_back(-1);
else if (M < static_cast<int>(NumElts)) {
ReconstructMask.push_back(FindIndex(V1, M));
} else {
ReconstructMask.push_back(NumElts + FindIndex(V2, M));
}
}
ReconstructMasks.push_back(std::move(ReconstructMask));
}
// Calculate the masks needed for the new input shuffles, which get padded
// with undef
SmallVector<int> V1A, V1B, V2A, V2B;
for (unsigned I = 0; I < V1.size(); I++) {
V1A.push_back(GetBaseMaskValue(SVI0A, V1[I].first));
V1B.push_back(GetBaseMaskValue(SVI0B, V1[I].first));
}
for (unsigned I = 0; I < V2.size(); I++) {
V2A.push_back(GetBaseMaskValue(SVI1A, V2[I].first));
V2B.push_back(GetBaseMaskValue(SVI1B, V2[I].first));
}
while (V1A.size() < NumElts) {
V1A.push_back(UndefMaskElem);
V1B.push_back(UndefMaskElem);
}
while (V2A.size() < NumElts) {
V2A.push_back(UndefMaskElem);
V2B.push_back(UndefMaskElem);
}
auto AddShuffleCost = [&](InstructionCost C, Instruction *I) {
auto *SV = dyn_cast<ShuffleVectorInst>(I);
if (!SV)
return C;
return C + TTI.getShuffleCost(isa<UndefValue>(SV->getOperand(1))
? TTI::SK_PermuteSingleSrc
: TTI::SK_PermuteTwoSrc,
VT, SV->getShuffleMask());
};
auto AddShuffleMaskCost = [&](InstructionCost C, ArrayRef<int> Mask) {
return C + TTI.getShuffleCost(TTI::SK_PermuteTwoSrc, VT, Mask);
};
// Get the costs of the shuffles + binops before and after with the new
// shuffle masks.
InstructionCost CostBefore =
TTI.getArithmeticInstrCost(Op0->getOpcode(), VT) +
TTI.getArithmeticInstrCost(Op1->getOpcode(), VT);
CostBefore += std::accumulate(Shuffles.begin(), Shuffles.end(),
InstructionCost(0), AddShuffleCost);
CostBefore += std::accumulate(InputShuffles.begin(), InputShuffles.end(),
InstructionCost(0), AddShuffleCost);
// The new binops will be unused for lanes past the used shuffle lengths.
// These types attempt to get the correct cost for that from the target.
FixedVectorType *Op0SmallVT =
FixedVectorType::get(VT->getScalarType(), V1.size());
FixedVectorType *Op1SmallVT =
FixedVectorType::get(VT->getScalarType(), V2.size());
InstructionCost CostAfter =
TTI.getArithmeticInstrCost(Op0->getOpcode(), Op0SmallVT) +
TTI.getArithmeticInstrCost(Op1->getOpcode(), Op1SmallVT);
CostAfter += std::accumulate(ReconstructMasks.begin(), ReconstructMasks.end(),
InstructionCost(0), AddShuffleMaskCost);
std::set<SmallVector<int>> OutputShuffleMasks({V1A, V1B, V2A, V2B});
CostAfter +=
std::accumulate(OutputShuffleMasks.begin(), OutputShuffleMasks.end(),
InstructionCost(0), AddShuffleMaskCost);
LLVM_DEBUG(dbgs() << "Found a binop select shuffle pattern: " << I << "\n");
LLVM_DEBUG(dbgs() << " CostBefore: " << CostBefore
<< " vs CostAfter: " << CostAfter << "\n");
if (CostBefore <= CostAfter)
return false;
// The cost model has passed, create the new instructions.
auto GetShuffleOperand = [&](Instruction *I, unsigned Op) -> Value * {
auto *SV = dyn_cast<ShuffleVectorInst>(I);
if (!SV)
return I;
if (isa<UndefValue>(SV->getOperand(1)))
if (auto *SSV = dyn_cast<ShuffleVectorInst>(SV->getOperand(0)))
if (InputShuffles.contains(SSV))
return SSV->getOperand(Op);
return SV->getOperand(Op);
};
Builder.SetInsertPoint(SVI0A->getNextNode());
Value *NSV0A = Builder.CreateShuffleVector(GetShuffleOperand(SVI0A, 0),
GetShuffleOperand(SVI0A, 1), V1A);
Builder.SetInsertPoint(SVI0B->getNextNode());
Value *NSV0B = Builder.CreateShuffleVector(GetShuffleOperand(SVI0B, 0),
GetShuffleOperand(SVI0B, 1), V1B);
Builder.SetInsertPoint(SVI1A->getNextNode());
Value *NSV1A = Builder.CreateShuffleVector(GetShuffleOperand(SVI1A, 0),
GetShuffleOperand(SVI1A, 1), V2A);
Builder.SetInsertPoint(SVI1B->getNextNode());
Value *NSV1B = Builder.CreateShuffleVector(GetShuffleOperand(SVI1B, 0),
GetShuffleOperand(SVI1B, 1), V2B);
Builder.SetInsertPoint(Op0);
Value *NOp0 = Builder.CreateBinOp((Instruction::BinaryOps)Op0->getOpcode(),
NSV0A, NSV0B);
if (auto *I = dyn_cast<Instruction>(NOp0))
I->copyIRFlags(Op0, true);
Builder.SetInsertPoint(Op1);
Value *NOp1 = Builder.CreateBinOp((Instruction::BinaryOps)Op1->getOpcode(),
NSV1A, NSV1B);
if (auto *I = dyn_cast<Instruction>(NOp1))
I->copyIRFlags(Op1, true);
for (int S = 0, E = ReconstructMasks.size(); S != E; S++) {
Builder.SetInsertPoint(Shuffles[S]);
Value *NSV = Builder.CreateShuffleVector(NOp0, NOp1, ReconstructMasks[S]);
replaceValue(*Shuffles[S], *NSV);
}
Worklist.pushValue(NSV0A);
Worklist.pushValue(NSV0B);
Worklist.pushValue(NSV1A);
Worklist.pushValue(NSV1B);
for (auto *S : Shuffles)
Worklist.add(S);
return true;
}
/// This is the entry point for all transforms. Pass manager differences are
/// handled in the callers of this function.
bool VectorCombine::run() {
if (DisableVectorCombine)
return false;
// Don't attempt vectorization if the target does not support vectors.
if (!TTI.getNumberOfRegisters(TTI.getRegisterClassForType(/*Vector*/ true)))
return false;
bool MadeChange = false;
auto FoldInst = [this, &MadeChange](Instruction &I) {
Builder.SetInsertPoint(&I);
if (!ScalarizationOnly) {
MadeChange |= vectorizeLoadInsert(I);
MadeChange |= foldExtractExtract(I);
MadeChange |= foldBitcastShuf(I);
MadeChange |= foldExtractedCmps(I);
MadeChange |= foldShuffleOfBinops(I);
MadeChange |= foldShuffleFromReductions(I);
MadeChange |= foldSelectShuffle(I);
}
MadeChange |= scalarizeBinopOrCmp(I);
MadeChange |= scalarizeLoadExtract(I);
MadeChange |= foldSingleElementStore(I);
};
for (BasicBlock &BB : F) {
// Ignore unreachable basic blocks.
if (!DT.isReachableFromEntry(&BB))
continue;
// Use early increment range so that we can erase instructions in loop.
for (Instruction &I : make_early_inc_range(BB)) {
if (I.isDebugOrPseudoInst())
continue;
FoldInst(I);
}
}
while (!Worklist.isEmpty()) {
Instruction *I = Worklist.removeOne();
if (!I)
continue;
if (isInstructionTriviallyDead(I)) {
eraseInstruction(*I);
continue;
}
FoldInst(*I);
}
return MadeChange;
}
// Pass manager boilerplate below here.
namespace {
class VectorCombineLegacyPass : public FunctionPass {
public:
static char ID;
VectorCombineLegacyPass() : FunctionPass(ID) {
initializeVectorCombineLegacyPassPass(*PassRegistry::getPassRegistry());
}
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.addRequired<AssumptionCacheTracker>();
AU.addRequired<DominatorTreeWrapperPass>();
AU.addRequired<TargetTransformInfoWrapperPass>();
AU.addRequired<AAResultsWrapperPass>();
AU.setPreservesCFG();
AU.addPreserved<DominatorTreeWrapperPass>();
AU.addPreserved<GlobalsAAWrapperPass>();
AU.addPreserved<AAResultsWrapperPass>();
AU.addPreserved<BasicAAWrapperPass>();
FunctionPass::getAnalysisUsage(AU);
}
bool runOnFunction(Function &F) override {
if (skipFunction(F))
return false;
auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
VectorCombine Combiner(F, TTI, DT, AA, AC, false);
return Combiner.run();
}
};
} // namespace
char VectorCombineLegacyPass::ID = 0;
INITIALIZE_PASS_BEGIN(VectorCombineLegacyPass, "vector-combine",
"Optimize scalar/vector ops", false,
false)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_END(VectorCombineLegacyPass, "vector-combine",
"Optimize scalar/vector ops", false, false)
Pass *llvm::createVectorCombinePass() {
return new VectorCombineLegacyPass();
}
PreservedAnalyses VectorCombinePass::run(Function &F,
FunctionAnalysisManager &FAM) {
auto &AC = FAM.getResult<AssumptionAnalysis>(F);
TargetTransformInfo &TTI = FAM.getResult<TargetIRAnalysis>(F);
DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(F);
AAResults &AA = FAM.getResult<AAManager>(F);
VectorCombine Combiner(F, TTI, DT, AA, AC, ScalarizationOnly);
if (!Combiner.run())
return PreservedAnalyses::all();
PreservedAnalyses PA;
PA.preserveSet<CFGAnalyses>();
return PA;
}
Reference in New Issue View Git Blame Copy Permalink