llvm/lib/CodeGen/InterleavedAccessPass.cpp
Eli Friedman 1dc854412e [InterleavedAccessPass] Don't increase the number of bytes loaded.
Even if the interleaving transform would otherwise be legal, we shouldn't
introduce an interleaved load that is wider than the original load: it might
have undefined behavior.

It might be possible to perform some sort of mask-narrowing transform in
some cases (using a narrower interleaved load, then extending the
results using shufflevectors).  But I haven't tried to implement that,
at least for now.

Fixes https://bugs.llvm.org/show_bug.cgi?id=41245 .

Differential Revision: https://reviews.llvm.org/D59954



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@357212 91177308-0d34-0410-b5e6-96231b3b80d8
2019-03-28 20:44:50 +00:00

473 lines
16 KiB
C++

//===- InterleavedAccessPass.cpp ------------------------------------------===//
//
// 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 file implements the Interleaved Access pass, which identifies
// interleaved memory accesses and transforms them into target specific
// intrinsics.
//
// An interleaved load reads data from memory into several vectors, with
// DE-interleaving the data on a factor. An interleaved store writes several
// vectors to memory with RE-interleaving the data on a factor.
//
// As interleaved accesses are difficult to identified in CodeGen (mainly
// because the VECTOR_SHUFFLE DAG node is quite different from the shufflevector
// IR), we identify and transform them to intrinsics in this pass so the
// intrinsics can be easily matched into target specific instructions later in
// CodeGen.
//
// E.g. An interleaved load (Factor = 2):
// %wide.vec = load <8 x i32>, <8 x i32>* %ptr
// %v0 = shuffle <8 x i32> %wide.vec, <8 x i32> undef, <0, 2, 4, 6>
// %v1 = shuffle <8 x i32> %wide.vec, <8 x i32> undef, <1, 3, 5, 7>
//
// It could be transformed into a ld2 intrinsic in AArch64 backend or a vld2
// intrinsic in ARM backend.
//
// In X86, this can be further optimized into a set of target
// specific loads followed by an optimized sequence of shuffles.
//
// E.g. An interleaved store (Factor = 3):
// %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
// <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
// store <12 x i32> %i.vec, <12 x i32>* %ptr
//
// It could be transformed into a st3 intrinsic in AArch64 backend or a vst3
// intrinsic in ARM backend.
//
// Similarly, a set of interleaved stores can be transformed into an optimized
// sequence of shuffles followed by a set of target specific stores for X86.
//
//===----------------------------------------------------------------------===//
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/CodeGen/TargetLowering.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Type.h"
#include "llvm/Pass.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetMachine.h"
#include <cassert>
#include <utility>
using namespace llvm;
#define DEBUG_TYPE "interleaved-access"
static cl::opt<bool> LowerInterleavedAccesses(
"lower-interleaved-accesses",
cl::desc("Enable lowering interleaved accesses to intrinsics"),
cl::init(true), cl::Hidden);
namespace {
class InterleavedAccess : public FunctionPass {
public:
static char ID;
InterleavedAccess() : FunctionPass(ID) {
initializeInterleavedAccessPass(*PassRegistry::getPassRegistry());
}
StringRef getPassName() const override { return "Interleaved Access Pass"; }
bool runOnFunction(Function &F) override;
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.addRequired<DominatorTreeWrapperPass>();
AU.addPreserved<DominatorTreeWrapperPass>();
}
private:
DominatorTree *DT = nullptr;
const TargetLowering *TLI = nullptr;
/// The maximum supported interleave factor.
unsigned MaxFactor;
/// Transform an interleaved load into target specific intrinsics.
bool lowerInterleavedLoad(LoadInst *LI,
SmallVector<Instruction *, 32> &DeadInsts);
/// Transform an interleaved store into target specific intrinsics.
bool lowerInterleavedStore(StoreInst *SI,
SmallVector<Instruction *, 32> &DeadInsts);
/// Returns true if the uses of an interleaved load by the
/// extractelement instructions in \p Extracts can be replaced by uses of the
/// shufflevector instructions in \p Shuffles instead. If so, the necessary
/// replacements are also performed.
bool tryReplaceExtracts(ArrayRef<ExtractElementInst *> Extracts,
ArrayRef<ShuffleVectorInst *> Shuffles);
};
} // end anonymous namespace.
char InterleavedAccess::ID = 0;
INITIALIZE_PASS_BEGIN(InterleavedAccess, DEBUG_TYPE,
"Lower interleaved memory accesses to target specific intrinsics", false,
false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_END(InterleavedAccess, DEBUG_TYPE,
"Lower interleaved memory accesses to target specific intrinsics", false,
false)
FunctionPass *llvm::createInterleavedAccessPass() {
return new InterleavedAccess();
}
/// Check if the mask is a DE-interleave mask of the given factor
/// \p Factor like:
/// <Index, Index+Factor, ..., Index+(NumElts-1)*Factor>
static bool isDeInterleaveMaskOfFactor(ArrayRef<int> Mask, unsigned Factor,
unsigned &Index) {
// Check all potential start indices from 0 to (Factor - 1).
for (Index = 0; Index < Factor; Index++) {
unsigned i = 0;
// Check that elements are in ascending order by Factor. Ignore undef
// elements.
for (; i < Mask.size(); i++)
if (Mask[i] >= 0 && static_cast<unsigned>(Mask[i]) != Index + i * Factor)
break;
if (i == Mask.size())
return true;
}
return false;
}
/// Check if the mask is a DE-interleave mask for an interleaved load.
///
/// E.g. DE-interleave masks (Factor = 2) could be:
/// <0, 2, 4, 6> (mask of index 0 to extract even elements)
/// <1, 3, 5, 7> (mask of index 1 to extract odd elements)
static bool isDeInterleaveMask(ArrayRef<int> Mask, unsigned &Factor,
unsigned &Index, unsigned MaxFactor,
unsigned NumLoadElements) {
if (Mask.size() < 2)
return false;
// Check potential Factors.
for (Factor = 2; Factor <= MaxFactor; Factor++) {
// Make sure we don't produce a load wider than the input load.
if (Mask.size() * Factor > NumLoadElements)
return false;
if (isDeInterleaveMaskOfFactor(Mask, Factor, Index))
return true;
}
return false;
}
/// Check if the mask can be used in an interleaved store.
//
/// It checks for a more general pattern than the RE-interleave mask.
/// I.e. <x, y, ... z, x+1, y+1, ...z+1, x+2, y+2, ...z+2, ...>
/// E.g. For a Factor of 2 (LaneLen=4): <4, 32, 5, 33, 6, 34, 7, 35>
/// E.g. For a Factor of 3 (LaneLen=4): <4, 32, 16, 5, 33, 17, 6, 34, 18, 7, 35, 19>
/// E.g. For a Factor of 4 (LaneLen=2): <8, 2, 12, 4, 9, 3, 13, 5>
///
/// The particular case of an RE-interleave mask is:
/// I.e. <0, LaneLen, ... , LaneLen*(Factor - 1), 1, LaneLen + 1, ...>
/// E.g. For a Factor of 2 (LaneLen=4): <0, 4, 1, 5, 2, 6, 3, 7>
static bool isReInterleaveMask(ArrayRef<int> Mask, unsigned &Factor,
unsigned MaxFactor, unsigned OpNumElts) {
unsigned NumElts = Mask.size();
if (NumElts < 4)
return false;
// Check potential Factors.
for (Factor = 2; Factor <= MaxFactor; Factor++) {
if (NumElts % Factor)
continue;
unsigned LaneLen = NumElts / Factor;
if (!isPowerOf2_32(LaneLen))
continue;
// Check whether each element matches the general interleaved rule.
// Ignore undef elements, as long as the defined elements match the rule.
// Outer loop processes all factors (x, y, z in the above example)
unsigned I = 0, J;
for (; I < Factor; I++) {
unsigned SavedLaneValue;
unsigned SavedNoUndefs = 0;
// Inner loop processes consecutive accesses (x, x+1... in the example)
for (J = 0; J < LaneLen - 1; J++) {
// Lane computes x's position in the Mask
unsigned Lane = J * Factor + I;
unsigned NextLane = Lane + Factor;
int LaneValue = Mask[Lane];
int NextLaneValue = Mask[NextLane];
// If both are defined, values must be sequential
if (LaneValue >= 0 && NextLaneValue >= 0 &&
LaneValue + 1 != NextLaneValue)
break;
// If the next value is undef, save the current one as reference
if (LaneValue >= 0 && NextLaneValue < 0) {
SavedLaneValue = LaneValue;
SavedNoUndefs = 1;
}
// Undefs are allowed, but defined elements must still be consecutive:
// i.e.: x,..., undef,..., x + 2,..., undef,..., undef,..., x + 5, ....
// Verify this by storing the last non-undef followed by an undef
// Check that following non-undef masks are incremented with the
// corresponding distance.
if (SavedNoUndefs > 0 && LaneValue < 0) {
SavedNoUndefs++;
if (NextLaneValue >= 0 &&
SavedLaneValue + SavedNoUndefs != (unsigned)NextLaneValue)
break;
}
}
if (J < LaneLen - 1)
break;
int StartMask = 0;
if (Mask[I] >= 0) {
// Check that the start of the I range (J=0) is greater than 0
StartMask = Mask[I];
} else if (Mask[(LaneLen - 1) * Factor + I] >= 0) {
// StartMask defined by the last value in lane
StartMask = Mask[(LaneLen - 1) * Factor + I] - J;
} else if (SavedNoUndefs > 0) {
// StartMask defined by some non-zero value in the j loop
StartMask = SavedLaneValue - (LaneLen - 1 - SavedNoUndefs);
}
// else StartMask remains set to 0, i.e. all elements are undefs
if (StartMask < 0)
break;
// We must stay within the vectors; This case can happen with undefs.
if (StartMask + LaneLen > OpNumElts*2)
break;
}
// Found an interleaved mask of current factor.
if (I == Factor)
return true;
}
return false;
}
bool InterleavedAccess::lowerInterleavedLoad(
LoadInst *LI, SmallVector<Instruction *, 32> &DeadInsts) {
if (!LI->isSimple())
return false;
SmallVector<ShuffleVectorInst *, 4> Shuffles;
SmallVector<ExtractElementInst *, 4> Extracts;
// Check if all users of this load are shufflevectors. If we encounter any
// users that are extractelement instructions, we save them to later check if
// they can be modifed to extract from one of the shufflevectors instead of
// the load.
for (auto UI = LI->user_begin(), E = LI->user_end(); UI != E; UI++) {
auto *Extract = dyn_cast<ExtractElementInst>(*UI);
if (Extract && isa<ConstantInt>(Extract->getIndexOperand())) {
Extracts.push_back(Extract);
continue;
}
ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(*UI);
if (!SVI || !isa<UndefValue>(SVI->getOperand(1)))
return false;
Shuffles.push_back(SVI);
}
if (Shuffles.empty())
return false;
unsigned Factor, Index;
unsigned NumLoadElements = LI->getType()->getVectorNumElements();
// Check if the first shufflevector is DE-interleave shuffle.
if (!isDeInterleaveMask(Shuffles[0]->getShuffleMask(), Factor, Index,
MaxFactor, NumLoadElements))
return false;
// Holds the corresponding index for each DE-interleave shuffle.
SmallVector<unsigned, 4> Indices;
Indices.push_back(Index);
Type *VecTy = Shuffles[0]->getType();
// Check if other shufflevectors are also DE-interleaved of the same type
// and factor as the first shufflevector.
for (unsigned i = 1; i < Shuffles.size(); i++) {
if (Shuffles[i]->getType() != VecTy)
return false;
if (!isDeInterleaveMaskOfFactor(Shuffles[i]->getShuffleMask(), Factor,
Index))
return false;
Indices.push_back(Index);
}
// Try and modify users of the load that are extractelement instructions to
// use the shufflevector instructions instead of the load.
if (!tryReplaceExtracts(Extracts, Shuffles))
return false;
LLVM_DEBUG(dbgs() << "IA: Found an interleaved load: " << *LI << "\n");
// Try to create target specific intrinsics to replace the load and shuffles.
if (!TLI->lowerInterleavedLoad(LI, Shuffles, Indices, Factor))
return false;
for (auto SVI : Shuffles)
DeadInsts.push_back(SVI);
DeadInsts.push_back(LI);
return true;
}
bool InterleavedAccess::tryReplaceExtracts(
ArrayRef<ExtractElementInst *> Extracts,
ArrayRef<ShuffleVectorInst *> Shuffles) {
// If there aren't any extractelement instructions to modify, there's nothing
// to do.
if (Extracts.empty())
return true;
// Maps extractelement instructions to vector-index pairs. The extractlement
// instructions will be modified to use the new vector and index operands.
DenseMap<ExtractElementInst *, std::pair<Value *, int>> ReplacementMap;
for (auto *Extract : Extracts) {
// The vector index that is extracted.
auto *IndexOperand = cast<ConstantInt>(Extract->getIndexOperand());
auto Index = IndexOperand->getSExtValue();
// Look for a suitable shufflevector instruction. The goal is to modify the
// extractelement instruction (which uses an interleaved load) to use one
// of the shufflevector instructions instead of the load.
for (auto *Shuffle : Shuffles) {
// If the shufflevector instruction doesn't dominate the extract, we
// can't create a use of it.
if (!DT->dominates(Shuffle, Extract))
continue;
// Inspect the indices of the shufflevector instruction. If the shuffle
// selects the same index that is extracted, we can modify the
// extractelement instruction.
SmallVector<int, 4> Indices;
Shuffle->getShuffleMask(Indices);
for (unsigned I = 0; I < Indices.size(); ++I)
if (Indices[I] == Index) {
assert(Extract->getOperand(0) == Shuffle->getOperand(0) &&
"Vector operations do not match");
ReplacementMap[Extract] = std::make_pair(Shuffle, I);
break;
}
// If we found a suitable shufflevector instruction, stop looking.
if (ReplacementMap.count(Extract))
break;
}
// If we did not find a suitable shufflevector instruction, the
// extractelement instruction cannot be modified, so we must give up.
if (!ReplacementMap.count(Extract))
return false;
}
// Finally, perform the replacements.
IRBuilder<> Builder(Extracts[0]->getContext());
for (auto &Replacement : ReplacementMap) {
auto *Extract = Replacement.first;
auto *Vector = Replacement.second.first;
auto Index = Replacement.second.second;
Builder.SetInsertPoint(Extract);
Extract->replaceAllUsesWith(Builder.CreateExtractElement(Vector, Index));
Extract->eraseFromParent();
}
return true;
}
bool InterleavedAccess::lowerInterleavedStore(
StoreInst *SI, SmallVector<Instruction *, 32> &DeadInsts) {
if (!SI->isSimple())
return false;
ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(SI->getValueOperand());
if (!SVI || !SVI->hasOneUse())
return false;
// Check if the shufflevector is RE-interleave shuffle.
unsigned Factor;
unsigned OpNumElts = SVI->getOperand(0)->getType()->getVectorNumElements();
if (!isReInterleaveMask(SVI->getShuffleMask(), Factor, MaxFactor, OpNumElts))
return false;
LLVM_DEBUG(dbgs() << "IA: Found an interleaved store: " << *SI << "\n");
// Try to create target specific intrinsics to replace the store and shuffle.
if (!TLI->lowerInterleavedStore(SI, SVI, Factor))
return false;
// Already have a new target specific interleaved store. Erase the old store.
DeadInsts.push_back(SI);
DeadInsts.push_back(SVI);
return true;
}
bool InterleavedAccess::runOnFunction(Function &F) {
auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
if (!TPC || !LowerInterleavedAccesses)
return false;
LLVM_DEBUG(dbgs() << "*** " << getPassName() << ": " << F.getName() << "\n");
DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
auto &TM = TPC->getTM<TargetMachine>();
TLI = TM.getSubtargetImpl(F)->getTargetLowering();
MaxFactor = TLI->getMaxSupportedInterleaveFactor();
// Holds dead instructions that will be erased later.
SmallVector<Instruction *, 32> DeadInsts;
bool Changed = false;
for (auto &I : instructions(F)) {
if (LoadInst *LI = dyn_cast<LoadInst>(&I))
Changed |= lowerInterleavedLoad(LI, DeadInsts);
if (StoreInst *SI = dyn_cast<StoreInst>(&I))
Changed |= lowerInterleavedStore(SI, DeadInsts);
}
for (auto I : DeadInsts)
I->eraseFromParent();
return Changed;
}