1 //===--- ExpandMemCmp.cpp - Expand memcmp() to load/stores ----------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This pass tries to expand memcmp() calls into optimally-sized loads and
10 // compares for the target.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/ADT/Statistic.h"
15 #include "llvm/Analysis/ConstantFolding.h"
16 #include "llvm/Analysis/DomTreeUpdater.h"
17 #include "llvm/Analysis/GlobalsModRef.h"
18 #include "llvm/Analysis/TargetLibraryInfo.h"
19 #include "llvm/Analysis/TargetTransformInfo.h"
20 #include "llvm/Analysis/ValueTracking.h"
21 #include "llvm/CodeGen/TargetSubtargetInfo.h"
22 #include "llvm/IR/Dominators.h"
23 #include "llvm/IR/IRBuilder.h"
24 #include "llvm/Transforms/Scalar.h"
28 #define DEBUG_TYPE "expandmemcmp"
30 STATISTIC(NumMemCmpCalls, "Number of memcmp calls");
31 STATISTIC(NumMemCmpNotConstant, "Number of memcmp calls without constant size");
32 STATISTIC(NumMemCmpGreaterThanMax,
33 "Number of memcmp calls with size greater than max size");
34 STATISTIC(NumMemCmpInlined, "Number of inlined memcmp calls");
36 static cl::opt<unsigned> MemCmpEqZeroNumLoadsPerBlock(
37 "memcmp-num-loads-per-block", cl::Hidden, cl::init(1),
38 cl::desc("The number of loads per basic block for inline expansion of "
39 "memcmp that is only being compared against zero."));
41 static cl::opt<unsigned> MaxLoadsPerMemcmp(
42 "max-loads-per-memcmp", cl::Hidden,
43 cl::desc("Set maximum number of loads used in expanded memcmp"));
45 static cl::opt<unsigned> MaxLoadsPerMemcmpOptSize(
46 "max-loads-per-memcmp-opt-size", cl::Hidden,
47 cl::desc("Set maximum number of loads used in expanded memcmp for -Os/Oz"));
51 // This class provides helper functions to expand a memcmp library call into an
53 class MemCmpExpansion {
55 BasicBlock *BB = nullptr;
56 PHINode *PhiSrc1 = nullptr;
57 PHINode *PhiSrc2 = nullptr;
59 ResultBlock() = default;
66 uint64_t NumLoadsNonOneByte;
67 const uint64_t NumLoadsPerBlockForZeroCmp;
68 std::vector<BasicBlock *> LoadCmpBlocks;
69 BasicBlock *EndBlock = nullptr;
71 const bool IsUsedForZeroCmp;
75 // Represents the decomposition in blocks of the expansion. For example,
76 // comparing 33 bytes on X86+sse can be done with 2x16-byte loads and
77 // 1x1-byte load, which would be represented as [{16, 0}, {16, 16}, {32, 1}.
79 LoadEntry(unsigned LoadSize, uint64_t Offset)
80 : LoadSize(LoadSize), Offset(Offset) {}
82 // The size of the load for this block, in bytes.
84 // The offset of this load from the base pointer, in bytes.
87 using LoadEntryVector = SmallVector<LoadEntry, 8>;
88 LoadEntryVector LoadSequence;
90 void createLoadCmpBlocks();
91 void createResultBlock();
92 void setupResultBlockPHINodes();
93 void setupEndBlockPHINodes();
94 Value *getCompareLoadPairs(unsigned BlockIndex, unsigned &LoadIndex);
95 void emitLoadCompareBlock(unsigned BlockIndex);
96 void emitLoadCompareBlockMultipleLoads(unsigned BlockIndex,
98 void emitLoadCompareByteBlock(unsigned BlockIndex, unsigned OffsetBytes);
99 void emitMemCmpResultBlock();
100 Value *getMemCmpExpansionZeroCase();
101 Value *getMemCmpEqZeroOneBlock();
102 Value *getMemCmpOneBlock();
103 Value *getPtrToElementAtOffset(Value *Source, Type *LoadSizeType,
104 uint64_t OffsetBytes);
106 static LoadEntryVector
107 computeGreedyLoadSequence(uint64_t Size, llvm::ArrayRef<unsigned> LoadSizes,
108 unsigned MaxNumLoads, unsigned &NumLoadsNonOneByte);
109 static LoadEntryVector
110 computeOverlappingLoadSequence(uint64_t Size, unsigned MaxLoadSize,
111 unsigned MaxNumLoads,
112 unsigned &NumLoadsNonOneByte);
115 MemCmpExpansion(CallInst *CI, uint64_t Size,
116 const TargetTransformInfo::MemCmpExpansionOptions &Options,
117 const bool IsUsedForZeroCmp, const DataLayout &TheDataLayout,
120 unsigned getNumBlocks();
121 uint64_t getNumLoads() const { return LoadSequence.size(); }
123 Value *getMemCmpExpansion();
126 MemCmpExpansion::LoadEntryVector MemCmpExpansion::computeGreedyLoadSequence(
127 uint64_t Size, llvm::ArrayRef<unsigned> LoadSizes,
128 const unsigned MaxNumLoads, unsigned &NumLoadsNonOneByte) {
129 NumLoadsNonOneByte = 0;
130 LoadEntryVector LoadSequence;
132 while (Size && !LoadSizes.empty()) {
133 const unsigned LoadSize = LoadSizes.front();
134 const uint64_t NumLoadsForThisSize = Size / LoadSize;
135 if (LoadSequence.size() + NumLoadsForThisSize > MaxNumLoads) {
136 // Do not expand if the total number of loads is larger than what the
137 // target allows. Note that it's important that we exit before completing
138 // the expansion to avoid using a ton of memory to store the expansion for
142 if (NumLoadsForThisSize > 0) {
143 for (uint64_t I = 0; I < NumLoadsForThisSize; ++I) {
144 LoadSequence.push_back({LoadSize, Offset});
148 ++NumLoadsNonOneByte;
149 Size = Size % LoadSize;
151 LoadSizes = LoadSizes.drop_front();
156 MemCmpExpansion::LoadEntryVector
157 MemCmpExpansion::computeOverlappingLoadSequence(uint64_t Size,
158 const unsigned MaxLoadSize,
159 const unsigned MaxNumLoads,
160 unsigned &NumLoadsNonOneByte) {
161 // These are already handled by the greedy approach.
162 if (Size < 2 || MaxLoadSize < 2)
165 // We try to do as many non-overlapping loads as possible starting from the
167 const uint64_t NumNonOverlappingLoads = Size / MaxLoadSize;
168 assert(NumNonOverlappingLoads && "there must be at least one load");
169 // There remain 0 to (MaxLoadSize - 1) bytes to load, this will be done with
170 // an overlapping load.
171 Size = Size - NumNonOverlappingLoads * MaxLoadSize;
172 // Bail if we do not need an overloapping store, this is already handled by
173 // the greedy approach.
176 // Bail if the number of loads (non-overlapping + potential overlapping one)
177 // is larger than the max allowed.
178 if ((NumNonOverlappingLoads + 1) > MaxNumLoads)
181 // Add non-overlapping loads.
182 LoadEntryVector LoadSequence;
184 for (uint64_t I = 0; I < NumNonOverlappingLoads; ++I) {
185 LoadSequence.push_back({MaxLoadSize, Offset});
186 Offset += MaxLoadSize;
189 // Add the last overlapping load.
190 assert(Size > 0 && Size < MaxLoadSize && "broken invariant");
191 LoadSequence.push_back({MaxLoadSize, Offset - (MaxLoadSize - Size)});
192 NumLoadsNonOneByte = 1;
196 // Initialize the basic block structure required for expansion of memcmp call
197 // with given maximum load size and memcmp size parameter.
198 // This structure includes:
199 // 1. A list of load compare blocks - LoadCmpBlocks.
200 // 2. An EndBlock, split from original instruction point, which is the block to
202 // 3. ResultBlock, block to branch to for early exit when a
203 // LoadCmpBlock finds a difference.
204 MemCmpExpansion::MemCmpExpansion(
205 CallInst *const CI, uint64_t Size,
206 const TargetTransformInfo::MemCmpExpansionOptions &Options,
207 const bool IsUsedForZeroCmp, const DataLayout &TheDataLayout,
209 : CI(CI), Size(Size), MaxLoadSize(0), NumLoadsNonOneByte(0),
210 NumLoadsPerBlockForZeroCmp(Options.NumLoadsPerBlock),
211 IsUsedForZeroCmp(IsUsedForZeroCmp), DL(TheDataLayout), Builder(CI),
212 DTU(DT, /*PostDominator*/ nullptr,
213 DomTreeUpdater::UpdateStrategy::Eager) {
214 assert(Size > 0 && "zero blocks");
215 // Scale the max size down if the target can load more bytes than we need.
216 llvm::ArrayRef<unsigned> LoadSizes(Options.LoadSizes);
217 while (!LoadSizes.empty() && LoadSizes.front() > Size) {
218 LoadSizes = LoadSizes.drop_front();
220 assert(!LoadSizes.empty() && "cannot load Size bytes");
221 MaxLoadSize = LoadSizes.front();
222 // Compute the decomposition.
223 unsigned GreedyNumLoadsNonOneByte = 0;
224 LoadSequence = computeGreedyLoadSequence(Size, LoadSizes, Options.MaxNumLoads,
225 GreedyNumLoadsNonOneByte);
226 NumLoadsNonOneByte = GreedyNumLoadsNonOneByte;
227 assert(LoadSequence.size() <= Options.MaxNumLoads && "broken invariant");
228 // If we allow overlapping loads and the load sequence is not already optimal,
229 // use overlapping loads.
230 if (Options.AllowOverlappingLoads &&
231 (LoadSequence.empty() || LoadSequence.size() > 2)) {
232 unsigned OverlappingNumLoadsNonOneByte = 0;
233 auto OverlappingLoads = computeOverlappingLoadSequence(
234 Size, MaxLoadSize, Options.MaxNumLoads, OverlappingNumLoadsNonOneByte);
235 if (!OverlappingLoads.empty() &&
236 (LoadSequence.empty() ||
237 OverlappingLoads.size() < LoadSequence.size())) {
238 LoadSequence = OverlappingLoads;
239 NumLoadsNonOneByte = OverlappingNumLoadsNonOneByte;
242 assert(LoadSequence.size() <= Options.MaxNumLoads && "broken invariant");
245 unsigned MemCmpExpansion::getNumBlocks() {
246 if (IsUsedForZeroCmp)
247 return getNumLoads() / NumLoadsPerBlockForZeroCmp +
248 (getNumLoads() % NumLoadsPerBlockForZeroCmp != 0 ? 1 : 0);
249 return getNumLoads();
252 void MemCmpExpansion::createLoadCmpBlocks() {
253 assert(ResBlock.BB && "ResBlock must be created before LoadCmpBlocks");
254 for (unsigned i = 0; i < getNumBlocks(); i++) {
255 BasicBlock *BB = BasicBlock::Create(CI->getContext(), "loadbb",
256 EndBlock->getParent(), EndBlock);
257 LoadCmpBlocks.push_back(BB);
261 void MemCmpExpansion::createResultBlock() {
262 assert(EndBlock && "EndBlock must be created before ResultBlock");
263 ResBlock.BB = BasicBlock::Create(CI->getContext(), "res_block",
264 EndBlock->getParent(), EndBlock);
267 /// Return a pointer to an element of type `LoadSizeType` at offset
269 Value *MemCmpExpansion::getPtrToElementAtOffset(Value *Source,
271 uint64_t OffsetBytes) {
272 if (OffsetBytes > 0) {
273 auto *ByteType = Type::getInt8Ty(CI->getContext());
274 Source = Builder.CreateGEP(
275 ByteType, Builder.CreateBitCast(Source, ByteType->getPointerTo()),
276 ConstantInt::get(ByteType, OffsetBytes));
278 return Builder.CreateBitCast(Source, LoadSizeType->getPointerTo());
281 // This function creates the IR instructions for loading and comparing 1 byte.
282 // It loads 1 byte from each source of the memcmp parameters with the given
283 // GEPIndex. It then subtracts the two loaded values and adds this result to the
284 // final phi node for selecting the memcmp result.
285 void MemCmpExpansion::emitLoadCompareByteBlock(unsigned BlockIndex,
286 unsigned OffsetBytes) {
287 BasicBlock *const BB = LoadCmpBlocks[BlockIndex];
288 Builder.SetInsertPoint(BB);
289 Type *LoadSizeType = Type::getInt8Ty(CI->getContext());
291 getPtrToElementAtOffset(CI->getArgOperand(0), LoadSizeType, OffsetBytes);
293 getPtrToElementAtOffset(CI->getArgOperand(1), LoadSizeType, OffsetBytes);
295 Value *LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
296 Value *LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
298 LoadSrc1 = Builder.CreateZExt(LoadSrc1, Type::getInt32Ty(CI->getContext()));
299 LoadSrc2 = Builder.CreateZExt(LoadSrc2, Type::getInt32Ty(CI->getContext()));
300 Value *Diff = Builder.CreateSub(LoadSrc1, LoadSrc2);
302 PhiRes->addIncoming(Diff, LoadCmpBlocks[BlockIndex]);
304 if (BlockIndex < (LoadCmpBlocks.size() - 1)) {
305 // Early exit branch if difference found to EndBlock. Otherwise, continue to
306 // next LoadCmpBlock,
307 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_NE, Diff,
308 ConstantInt::get(Diff->getType(), 0));
309 BasicBlock *const NextBB = LoadCmpBlocks[BlockIndex + 1];
310 BranchInst *CmpBr = BranchInst::Create(EndBlock, NextBB, Cmp);
311 Builder.Insert(CmpBr);
312 DTU.applyUpdates({{DominatorTree::Insert, BB, EndBlock},
313 {DominatorTree::Insert, BB, NextBB}});
315 // The last block has an unconditional branch to EndBlock.
316 BranchInst *CmpBr = BranchInst::Create(EndBlock);
317 Builder.Insert(CmpBr);
318 DTU.applyUpdates({{DominatorTree::Insert, BB, EndBlock}});
322 /// Generate an equality comparison for one or more pairs of loaded values.
323 /// This is used in the case where the memcmp() call is compared equal or not
325 Value *MemCmpExpansion::getCompareLoadPairs(unsigned BlockIndex,
326 unsigned &LoadIndex) {
327 assert(LoadIndex < getNumLoads() &&
328 "getCompareLoadPairs() called with no remaining loads");
329 std::vector<Value *> XorList, OrList;
330 Value *Diff = nullptr;
332 const unsigned NumLoads =
333 std::min(getNumLoads() - LoadIndex, NumLoadsPerBlockForZeroCmp);
335 // For a single-block expansion, start inserting before the memcmp call.
336 if (LoadCmpBlocks.empty())
337 Builder.SetInsertPoint(CI);
339 Builder.SetInsertPoint(LoadCmpBlocks[BlockIndex]);
341 Value *Cmp = nullptr;
342 // If we have multiple loads per block, we need to generate a composite
343 // comparison using xor+or. The type for the combinations is the largest load
345 IntegerType *const MaxLoadType =
346 NumLoads == 1 ? nullptr
347 : IntegerType::get(CI->getContext(), MaxLoadSize * 8);
348 for (unsigned i = 0; i < NumLoads; ++i, ++LoadIndex) {
349 const LoadEntry &CurLoadEntry = LoadSequence[LoadIndex];
351 IntegerType *LoadSizeType =
352 IntegerType::get(CI->getContext(), CurLoadEntry.LoadSize * 8);
354 Value *Source1 = getPtrToElementAtOffset(CI->getArgOperand(0), LoadSizeType,
355 CurLoadEntry.Offset);
356 Value *Source2 = getPtrToElementAtOffset(CI->getArgOperand(1), LoadSizeType,
357 CurLoadEntry.Offset);
359 // Get a constant or load a value for each source address.
360 Value *LoadSrc1 = nullptr;
361 if (auto *Source1C = dyn_cast<Constant>(Source1))
362 LoadSrc1 = ConstantFoldLoadFromConstPtr(Source1C, LoadSizeType, DL);
364 LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
366 Value *LoadSrc2 = nullptr;
367 if (auto *Source2C = dyn_cast<Constant>(Source2))
368 LoadSrc2 = ConstantFoldLoadFromConstPtr(Source2C, LoadSizeType, DL);
370 LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
373 if (LoadSizeType != MaxLoadType) {
374 LoadSrc1 = Builder.CreateZExt(LoadSrc1, MaxLoadType);
375 LoadSrc2 = Builder.CreateZExt(LoadSrc2, MaxLoadType);
377 // If we have multiple loads per block, we need to generate a composite
378 // comparison using xor+or.
379 Diff = Builder.CreateXor(LoadSrc1, LoadSrc2);
380 Diff = Builder.CreateZExt(Diff, MaxLoadType);
381 XorList.push_back(Diff);
383 // If there's only one load per block, we just compare the loaded values.
384 Cmp = Builder.CreateICmpNE(LoadSrc1, LoadSrc2);
388 auto pairWiseOr = [&](std::vector<Value *> &InList) -> std::vector<Value *> {
389 std::vector<Value *> OutList;
390 for (unsigned i = 0; i < InList.size() - 1; i = i + 2) {
391 Value *Or = Builder.CreateOr(InList[i], InList[i + 1]);
392 OutList.push_back(Or);
394 if (InList.size() % 2 != 0)
395 OutList.push_back(InList.back());
400 // Pairwise OR the XOR results.
401 OrList = pairWiseOr(XorList);
403 // Pairwise OR the OR results until one result left.
404 while (OrList.size() != 1) {
405 OrList = pairWiseOr(OrList);
408 assert(Diff && "Failed to find comparison diff");
409 Cmp = Builder.CreateICmpNE(OrList[0], ConstantInt::get(Diff->getType(), 0));
415 void MemCmpExpansion::emitLoadCompareBlockMultipleLoads(unsigned BlockIndex,
416 unsigned &LoadIndex) {
417 Value *Cmp = getCompareLoadPairs(BlockIndex, LoadIndex);
419 BasicBlock *NextBB = (BlockIndex == (LoadCmpBlocks.size() - 1))
421 : LoadCmpBlocks[BlockIndex + 1];
422 // Early exit branch if difference found to ResultBlock. Otherwise,
423 // continue to next LoadCmpBlock or EndBlock.
424 BranchInst *CmpBr = BranchInst::Create(ResBlock.BB, NextBB, Cmp);
425 Builder.Insert(CmpBr);
426 BasicBlock *const BB = LoadCmpBlocks[BlockIndex];
428 // Add a phi edge for the last LoadCmpBlock to Endblock with a value of 0
429 // since early exit to ResultBlock was not taken (no difference was found in
430 // any of the bytes).
431 if (BlockIndex == LoadCmpBlocks.size() - 1) {
432 Value *Zero = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 0);
433 PhiRes->addIncoming(Zero, BB);
435 DTU.applyUpdates({{DominatorTree::Insert, BB, ResBlock.BB},
436 {DominatorTree::Insert, BB, NextBB}});
439 // This function creates the IR intructions for loading and comparing using the
440 // given LoadSize. It loads the number of bytes specified by LoadSize from each
441 // source of the memcmp parameters. It then does a subtract to see if there was
442 // a difference in the loaded values. If a difference is found, it branches
443 // with an early exit to the ResultBlock for calculating which source was
444 // larger. Otherwise, it falls through to the either the next LoadCmpBlock or
445 // the EndBlock if this is the last LoadCmpBlock. Loading 1 byte is handled with
446 // a special case through emitLoadCompareByteBlock. The special handling can
447 // simply subtract the loaded values and add it to the result phi node.
448 void MemCmpExpansion::emitLoadCompareBlock(unsigned BlockIndex) {
449 // There is one load per block in this case, BlockIndex == LoadIndex.
450 const LoadEntry &CurLoadEntry = LoadSequence[BlockIndex];
452 if (CurLoadEntry.LoadSize == 1) {
453 MemCmpExpansion::emitLoadCompareByteBlock(BlockIndex, CurLoadEntry.Offset);
458 IntegerType::get(CI->getContext(), CurLoadEntry.LoadSize * 8);
459 Type *MaxLoadType = IntegerType::get(CI->getContext(), MaxLoadSize * 8);
460 assert(CurLoadEntry.LoadSize <= MaxLoadSize && "Unexpected load type");
462 BasicBlock *const BB = LoadCmpBlocks[BlockIndex];
463 Builder.SetInsertPoint(BB);
465 Value *Source1 = getPtrToElementAtOffset(CI->getArgOperand(0), LoadSizeType,
466 CurLoadEntry.Offset);
467 Value *Source2 = getPtrToElementAtOffset(CI->getArgOperand(1), LoadSizeType,
468 CurLoadEntry.Offset);
470 // Load LoadSizeType from the base address.
471 Value *LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
472 Value *LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
474 if (DL.isLittleEndian()) {
475 Function *Bswap = Intrinsic::getDeclaration(CI->getModule(),
476 Intrinsic::bswap, LoadSizeType);
477 LoadSrc1 = Builder.CreateCall(Bswap, LoadSrc1);
478 LoadSrc2 = Builder.CreateCall(Bswap, LoadSrc2);
481 if (LoadSizeType != MaxLoadType) {
482 LoadSrc1 = Builder.CreateZExt(LoadSrc1, MaxLoadType);
483 LoadSrc2 = Builder.CreateZExt(LoadSrc2, MaxLoadType);
486 // Add the loaded values to the phi nodes for calculating memcmp result only
487 // if result is not used in a zero equality.
488 if (!IsUsedForZeroCmp) {
489 ResBlock.PhiSrc1->addIncoming(LoadSrc1, LoadCmpBlocks[BlockIndex]);
490 ResBlock.PhiSrc2->addIncoming(LoadSrc2, LoadCmpBlocks[BlockIndex]);
493 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, LoadSrc1, LoadSrc2);
494 BasicBlock *NextBB = (BlockIndex == (LoadCmpBlocks.size() - 1))
496 : LoadCmpBlocks[BlockIndex + 1];
497 // Early exit branch if difference found to ResultBlock. Otherwise, continue
498 // to next LoadCmpBlock or EndBlock.
499 BranchInst *CmpBr = BranchInst::Create(NextBB, ResBlock.BB, Cmp);
500 Builder.Insert(CmpBr);
502 // Add a phi edge for the last LoadCmpBlock to Endblock with a value of 0
503 // since early exit to ResultBlock was not taken (no difference was found in
504 // any of the bytes).
505 if (BlockIndex == LoadCmpBlocks.size() - 1) {
506 Value *Zero = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 0);
507 PhiRes->addIncoming(Zero, BB);
509 DTU.applyUpdates({{DominatorTree::Insert, BB, ResBlock.BB},
510 {DominatorTree::Insert, BB, NextBB}});
513 // This function populates the ResultBlock with a sequence to calculate the
514 // memcmp result. It compares the two loaded source values and returns -1 if
515 // src1 < src2 and 1 if src1 > src2.
516 void MemCmpExpansion::emitMemCmpResultBlock() {
517 // Special case: if memcmp result is used in a zero equality, result does not
518 // need to be calculated and can simply return 1.
519 if (IsUsedForZeroCmp) {
520 BasicBlock::iterator InsertPt = ResBlock.BB->getFirstInsertionPt();
521 Builder.SetInsertPoint(ResBlock.BB, InsertPt);
522 Value *Res = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 1);
523 PhiRes->addIncoming(Res, ResBlock.BB);
524 BranchInst *NewBr = BranchInst::Create(EndBlock);
525 Builder.Insert(NewBr);
526 DTU.applyUpdates({{DominatorTree::Insert, ResBlock.BB, EndBlock}});
529 BasicBlock::iterator InsertPt = ResBlock.BB->getFirstInsertionPt();
530 Builder.SetInsertPoint(ResBlock.BB, InsertPt);
532 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_ULT, ResBlock.PhiSrc1,
536 Builder.CreateSelect(Cmp, ConstantInt::get(Builder.getInt32Ty(), -1),
537 ConstantInt::get(Builder.getInt32Ty(), 1));
539 BranchInst *NewBr = BranchInst::Create(EndBlock);
540 Builder.Insert(NewBr);
541 PhiRes->addIncoming(Res, ResBlock.BB);
542 DTU.applyUpdates({{DominatorTree::Insert, ResBlock.BB, EndBlock}});
545 void MemCmpExpansion::setupResultBlockPHINodes() {
546 Type *MaxLoadType = IntegerType::get(CI->getContext(), MaxLoadSize * 8);
547 Builder.SetInsertPoint(ResBlock.BB);
548 // Note: this assumes one load per block.
550 Builder.CreatePHI(MaxLoadType, NumLoadsNonOneByte, "phi.src1");
552 Builder.CreatePHI(MaxLoadType, NumLoadsNonOneByte, "phi.src2");
555 void MemCmpExpansion::setupEndBlockPHINodes() {
556 Builder.SetInsertPoint(&EndBlock->front());
557 PhiRes = Builder.CreatePHI(Type::getInt32Ty(CI->getContext()), 2, "phi.res");
560 Value *MemCmpExpansion::getMemCmpExpansionZeroCase() {
561 unsigned LoadIndex = 0;
562 // This loop populates each of the LoadCmpBlocks with the IR sequence to
563 // handle multiple loads per block.
564 for (unsigned I = 0; I < getNumBlocks(); ++I) {
565 emitLoadCompareBlockMultipleLoads(I, LoadIndex);
568 emitMemCmpResultBlock();
572 /// A memcmp expansion that compares equality with 0 and only has one block of
573 /// load and compare can bypass the compare, branch, and phi IR that is required
574 /// in the general case.
575 Value *MemCmpExpansion::getMemCmpEqZeroOneBlock() {
576 unsigned LoadIndex = 0;
577 Value *Cmp = getCompareLoadPairs(0, LoadIndex);
578 assert(LoadIndex == getNumLoads() && "some entries were not consumed");
579 return Builder.CreateZExt(Cmp, Type::getInt32Ty(CI->getContext()));
582 /// A memcmp expansion that only has one block of load and compare can bypass
583 /// the compare, branch, and phi IR that is required in the general case.
584 Value *MemCmpExpansion::getMemCmpOneBlock() {
585 Type *LoadSizeType = IntegerType::get(CI->getContext(), Size * 8);
586 Value *Source1 = CI->getArgOperand(0);
587 Value *Source2 = CI->getArgOperand(1);
589 // Cast source to LoadSizeType*.
590 if (Source1->getType() != LoadSizeType)
591 Source1 = Builder.CreateBitCast(Source1, LoadSizeType->getPointerTo());
592 if (Source2->getType() != LoadSizeType)
593 Source2 = Builder.CreateBitCast(Source2, LoadSizeType->getPointerTo());
595 // Load LoadSizeType from the base address.
596 Value *LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
597 Value *LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
599 if (DL.isLittleEndian() && Size != 1) {
600 Function *Bswap = Intrinsic::getDeclaration(CI->getModule(),
601 Intrinsic::bswap, LoadSizeType);
602 LoadSrc1 = Builder.CreateCall(Bswap, LoadSrc1);
603 LoadSrc2 = Builder.CreateCall(Bswap, LoadSrc2);
607 // The i8 and i16 cases don't need compares. We zext the loaded values and
608 // subtract them to get the suitable negative, zero, or positive i32 result.
609 LoadSrc1 = Builder.CreateZExt(LoadSrc1, Builder.getInt32Ty());
610 LoadSrc2 = Builder.CreateZExt(LoadSrc2, Builder.getInt32Ty());
611 return Builder.CreateSub(LoadSrc1, LoadSrc2);
614 // The result of memcmp is negative, zero, or positive, so produce that by
615 // subtracting 2 extended compare bits: sub (ugt, ult).
616 // If a target prefers to use selects to get -1/0/1, they should be able
617 // to transform this later. The inverse transform (going from selects to math)
618 // may not be possible in the DAG because the selects got converted into
619 // branches before we got there.
620 Value *CmpUGT = Builder.CreateICmpUGT(LoadSrc1, LoadSrc2);
621 Value *CmpULT = Builder.CreateICmpULT(LoadSrc1, LoadSrc2);
622 Value *ZextUGT = Builder.CreateZExt(CmpUGT, Builder.getInt32Ty());
623 Value *ZextULT = Builder.CreateZExt(CmpULT, Builder.getInt32Ty());
624 return Builder.CreateSub(ZextUGT, ZextULT);
627 // This function expands the memcmp call into an inline expansion and returns
628 // the memcmp result.
629 Value *MemCmpExpansion::getMemCmpExpansion() {
630 // Create the basic block framework for a multi-block expansion.
631 if (getNumBlocks() != 1) {
632 BasicBlock *StartBlock = CI->getParent();
633 EndBlock = StartBlock->splitBasicBlock(CI, "endblock");
634 DTU.applyUpdates({{DominatorTree::Insert, StartBlock, EndBlock}});
635 setupEndBlockPHINodes();
638 // If return value of memcmp is not used in a zero equality, we need to
639 // calculate which source was larger. The calculation requires the
640 // two loaded source values of each load compare block.
641 // These will be saved in the phi nodes created by setupResultBlockPHINodes.
642 if (!IsUsedForZeroCmp)
643 setupResultBlockPHINodes();
645 // Create the number of required load compare basic blocks.
646 createLoadCmpBlocks();
648 // Update the terminator added by splitBasicBlock to branch to the first
650 BasicBlock *const FirstLoadBB = LoadCmpBlocks[0];
651 StartBlock->getTerminator()->setSuccessor(0, FirstLoadBB);
652 DTU.applyUpdates({{DominatorTree::Delete, StartBlock, EndBlock},
653 {DominatorTree::Insert, StartBlock, FirstLoadBB}});
656 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
658 if (IsUsedForZeroCmp)
659 return getNumBlocks() == 1 ? getMemCmpEqZeroOneBlock()
660 : getMemCmpExpansionZeroCase();
662 if (getNumBlocks() == 1)
663 return getMemCmpOneBlock();
665 for (unsigned I = 0; I < getNumBlocks(); ++I) {
666 emitLoadCompareBlock(I);
669 emitMemCmpResultBlock();
673 // This function checks to see if an expansion of memcmp can be generated.
674 // It checks for constant compare size that is less than the max inline size.
675 // If an expansion cannot occur, returns false to leave as a library call.
676 // Otherwise, the library call is replaced with a new IR instruction sequence.
677 /// We want to transform:
678 /// %call = call signext i32 @memcmp(i8* %0, i8* %1, i64 15)
681 /// %0 = bitcast i32* %buffer2 to i8*
682 /// %1 = bitcast i32* %buffer1 to i8*
683 /// %2 = bitcast i8* %1 to i64*
684 /// %3 = bitcast i8* %0 to i64*
685 /// %4 = load i64, i64* %2
686 /// %5 = load i64, i64* %3
687 /// %6 = call i64 @llvm.bswap.i64(i64 %4)
688 /// %7 = call i64 @llvm.bswap.i64(i64 %5)
689 /// %8 = sub i64 %6, %7
690 /// %9 = icmp ne i64 %8, 0
691 /// br i1 %9, label %res_block, label %loadbb1
692 /// res_block: ; preds = %loadbb2,
693 /// %loadbb1, %loadbb
694 /// %phi.src1 = phi i64 [ %6, %loadbb ], [ %22, %loadbb1 ], [ %36, %loadbb2 ]
695 /// %phi.src2 = phi i64 [ %7, %loadbb ], [ %23, %loadbb1 ], [ %37, %loadbb2 ]
696 /// %10 = icmp ult i64 %phi.src1, %phi.src2
697 /// %11 = select i1 %10, i32 -1, i32 1
698 /// br label %endblock
699 /// loadbb1: ; preds = %loadbb
700 /// %12 = bitcast i32* %buffer2 to i8*
701 /// %13 = bitcast i32* %buffer1 to i8*
702 /// %14 = bitcast i8* %13 to i32*
703 /// %15 = bitcast i8* %12 to i32*
704 /// %16 = getelementptr i32, i32* %14, i32 2
705 /// %17 = getelementptr i32, i32* %15, i32 2
706 /// %18 = load i32, i32* %16
707 /// %19 = load i32, i32* %17
708 /// %20 = call i32 @llvm.bswap.i32(i32 %18)
709 /// %21 = call i32 @llvm.bswap.i32(i32 %19)
710 /// %22 = zext i32 %20 to i64
711 /// %23 = zext i32 %21 to i64
712 /// %24 = sub i64 %22, %23
713 /// %25 = icmp ne i64 %24, 0
714 /// br i1 %25, label %res_block, label %loadbb2
715 /// loadbb2: ; preds = %loadbb1
716 /// %26 = bitcast i32* %buffer2 to i8*
717 /// %27 = bitcast i32* %buffer1 to i8*
718 /// %28 = bitcast i8* %27 to i16*
719 /// %29 = bitcast i8* %26 to i16*
720 /// %30 = getelementptr i16, i16* %28, i16 6
721 /// %31 = getelementptr i16, i16* %29, i16 6
722 /// %32 = load i16, i16* %30
723 /// %33 = load i16, i16* %31
724 /// %34 = call i16 @llvm.bswap.i16(i16 %32)
725 /// %35 = call i16 @llvm.bswap.i16(i16 %33)
726 /// %36 = zext i16 %34 to i64
727 /// %37 = zext i16 %35 to i64
728 /// %38 = sub i64 %36, %37
729 /// %39 = icmp ne i64 %38, 0
730 /// br i1 %39, label %res_block, label %loadbb3
731 /// loadbb3: ; preds = %loadbb2
732 /// %40 = bitcast i32* %buffer2 to i8*
733 /// %41 = bitcast i32* %buffer1 to i8*
734 /// %42 = getelementptr i8, i8* %41, i8 14
735 /// %43 = getelementptr i8, i8* %40, i8 14
736 /// %44 = load i8, i8* %42
737 /// %45 = load i8, i8* %43
738 /// %46 = zext i8 %44 to i32
739 /// %47 = zext i8 %45 to i32
740 /// %48 = sub i32 %46, %47
741 /// br label %endblock
742 /// endblock: ; preds = %res_block,
744 /// %phi.res = phi i32 [ %48, %loadbb3 ], [ %11, %res_block ]
746 static bool expandMemCmp(CallInst *CI, const TargetTransformInfo *TTI,
747 const DataLayout *DL, DominatorTree *DT) {
750 // Early exit from expansion if -Oz.
751 if (CI->getFunction()->hasMinSize())
754 // Early exit from expansion if size is not a constant.
755 ConstantInt *SizeCast = dyn_cast<ConstantInt>(CI->getArgOperand(2));
757 NumMemCmpNotConstant++;
760 const uint64_t SizeVal = SizeCast->getZExtValue();
765 // TTI call to check if target would like to expand memcmp. Also, get the
766 // available load sizes.
767 const bool IsUsedForZeroCmp = isOnlyUsedInZeroEqualityComparison(CI);
768 auto Options = TTI->enableMemCmpExpansion(CI->getFunction()->hasOptSize(),
773 if (MemCmpEqZeroNumLoadsPerBlock.getNumOccurrences())
774 Options.NumLoadsPerBlock = MemCmpEqZeroNumLoadsPerBlock;
776 if (CI->getFunction()->hasOptSize() &&
777 MaxLoadsPerMemcmpOptSize.getNumOccurrences())
778 Options.MaxNumLoads = MaxLoadsPerMemcmpOptSize;
780 if (!CI->getFunction()->hasOptSize() && MaxLoadsPerMemcmp.getNumOccurrences())
781 Options.MaxNumLoads = MaxLoadsPerMemcmp;
783 MemCmpExpansion Expansion(CI, SizeVal, Options, IsUsedForZeroCmp, *DL, DT);
785 // Don't expand if this will require more loads than desired by the target.
786 if (Expansion.getNumLoads() == 0) {
787 NumMemCmpGreaterThanMax++;
793 Value *Res = Expansion.getMemCmpExpansion();
795 // Replace call with result of expansion and erase call.
796 CI->replaceAllUsesWith(Res);
797 CI->eraseFromParent();
802 class ExpandMemCmpPass : public FunctionPass {
806 ExpandMemCmpPass() : FunctionPass(ID) {
807 initializeExpandMemCmpPassPass(*PassRegistry::getPassRegistry());
810 bool runOnFunction(Function &F) override {
814 const TargetLibraryInfo *TLI =
815 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
816 const TargetTransformInfo *TTI =
817 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
818 // ExpandMemCmp does not need the DominatorTree, but we update it if it's
819 // already available.
820 auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
821 auto PA = runImpl(F, TLI, TTI, DTWP ? &DTWP->getDomTree() : nullptr);
822 return !PA.areAllPreserved();
826 void getAnalysisUsage(AnalysisUsage &AU) const override {
827 AU.addRequired<TargetLibraryInfoWrapperPass>();
828 AU.addRequired<TargetTransformInfoWrapperPass>();
829 AU.addUsedIfAvailable<DominatorTreeWrapperPass>();
830 AU.addPreserved<GlobalsAAWrapperPass>();
831 AU.addPreserved<DominatorTreeWrapperPass>();
832 FunctionPass::getAnalysisUsage(AU);
835 PreservedAnalyses runImpl(Function &F, const TargetLibraryInfo *TLI,
836 const TargetTransformInfo *TTI, DominatorTree *DT);
837 // Returns true if a change was made.
838 bool runOnBlock(BasicBlock &BB, const TargetLibraryInfo *TLI,
839 const TargetTransformInfo *TTI, const DataLayout &DL,
843 bool ExpandMemCmpPass::runOnBlock(BasicBlock &BB, const TargetLibraryInfo *TLI,
844 const TargetTransformInfo *TTI,
845 const DataLayout &DL, DominatorTree *DT) {
846 for (Instruction &I : BB) {
847 CallInst *CI = dyn_cast<CallInst>(&I);
852 if (TLI->getLibFunc(ImmutableCallSite(CI), Func) &&
853 (Func == LibFunc_memcmp || Func == LibFunc_bcmp) &&
854 expandMemCmp(CI, TTI, &DL, DT)) {
861 PreservedAnalyses ExpandMemCmpPass::runImpl(Function &F,
862 const TargetLibraryInfo *TLI,
863 const TargetTransformInfo *TTI,
865 const DataLayout &DL = F.getParent()->getDataLayout();
866 bool MadeChanges = false;
867 for (auto BBIt = F.begin(); BBIt != F.end();) {
868 if (runOnBlock(*BBIt, TLI, TTI, DL, DT)) {
870 // If changes were made, restart the function from the beginning, since
871 // the structure of the function was changed.
878 return PreservedAnalyses::all();
879 PreservedAnalyses PA;
880 PA.preserve<GlobalsAA>();
881 PA.preserve<DominatorTreeAnalysis>();
887 char ExpandMemCmpPass::ID = 0;
888 INITIALIZE_PASS_BEGIN(ExpandMemCmpPass, "expandmemcmp",
889 "Expand memcmp() to load/stores", false, false)
890 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
891 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
892 INITIALIZE_PASS_END(ExpandMemCmpPass, "expandmemcmp",
893 "Expand memcmp() to load/stores", false, false)
895 Pass *llvm::createExpandMemCmpPass() { return new ExpandMemCmpPass(); }