LLVM  9.0.0svn
ExpandMemCmp.cpp
Go to the documentation of this file.
1 //===--- ExpandMemCmp.cpp - Expand memcmp() to load/stores ----------------===//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass tries to expand memcmp() calls into optimally-sized loads and
10 // compares for the target.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/ADT/Statistic.h"
22 #include "llvm/IR/IRBuilder.h"
23 
24 using namespace llvm;
25 
26 #define DEBUG_TYPE "expandmemcmp"
27 
28 STATISTIC(NumMemCmpCalls, "Number of memcmp calls");
29 STATISTIC(NumMemCmpNotConstant, "Number of memcmp calls without constant size");
30 STATISTIC(NumMemCmpGreaterThanMax,
31  "Number of memcmp calls with size greater than max size");
32 STATISTIC(NumMemCmpInlined, "Number of inlined memcmp calls");
33 
35  "memcmp-num-loads-per-block", cl::Hidden, cl::init(1),
36  cl::desc("The number of loads per basic block for inline expansion of "
37  "memcmp that is only being compared against zero."));
38 
39 namespace {
40 
41 
42 // This class provides helper functions to expand a memcmp library call into an
43 // inline expansion.
44 class MemCmpExpansion {
45  struct ResultBlock {
46  BasicBlock *BB = nullptr;
47  PHINode *PhiSrc1 = nullptr;
48  PHINode *PhiSrc2 = nullptr;
49 
50  ResultBlock() = default;
51  };
52 
53  CallInst *const CI;
54  ResultBlock ResBlock;
55  const uint64_t Size;
56  unsigned MaxLoadSize;
57  uint64_t NumLoadsNonOneByte;
58  const uint64_t NumLoadsPerBlockForZeroCmp;
59  std::vector<BasicBlock *> LoadCmpBlocks;
60  BasicBlock *EndBlock;
61  PHINode *PhiRes;
62  const bool IsUsedForZeroCmp;
63  const DataLayout &DL;
64  IRBuilder<> Builder;
65  // Represents the decomposition in blocks of the expansion. For example,
66  // comparing 33 bytes on X86+sse can be done with 2x16-byte loads and
67  // 1x1-byte load, which would be represented as [{16, 0}, {16, 16}, {32, 1}.
68  struct LoadEntry {
69  LoadEntry(unsigned LoadSize, uint64_t Offset)
70  : LoadSize(LoadSize), Offset(Offset) {
71  }
72 
73  // The size of the load for this block, in bytes.
74  unsigned LoadSize;
75  // The offset of this load from the base pointer, in bytes.
76  uint64_t Offset;
77  };
78  using LoadEntryVector = SmallVector<LoadEntry, 8>;
79  LoadEntryVector LoadSequence;
80 
81  void createLoadCmpBlocks();
82  void createResultBlock();
83  void setupResultBlockPHINodes();
84  void setupEndBlockPHINodes();
85  Value *getCompareLoadPairs(unsigned BlockIndex, unsigned &LoadIndex);
86  void emitLoadCompareBlock(unsigned BlockIndex);
87  void emitLoadCompareBlockMultipleLoads(unsigned BlockIndex,
88  unsigned &LoadIndex);
89  void emitLoadCompareByteBlock(unsigned BlockIndex, unsigned OffsetBytes);
90  void emitMemCmpResultBlock();
91  Value *getMemCmpExpansionZeroCase();
92  Value *getMemCmpEqZeroOneBlock();
93  Value *getMemCmpOneBlock();
94  Value *getPtrToElementAtOffset(Value *Source, Type *LoadSizeType,
95  uint64_t OffsetBytes);
96 
97  static LoadEntryVector
98  computeGreedyLoadSequence(uint64_t Size, llvm::ArrayRef<unsigned> LoadSizes,
99  unsigned MaxNumLoads, unsigned &NumLoadsNonOneByte);
100  static LoadEntryVector
101  computeOverlappingLoadSequence(uint64_t Size, unsigned MaxLoadSize,
102  unsigned MaxNumLoads,
103  unsigned &NumLoadsNonOneByte);
104 
105 public:
106  MemCmpExpansion(CallInst *CI, uint64_t Size,
108  unsigned MaxNumLoads, const bool IsUsedForZeroCmp,
109  unsigned MaxLoadsPerBlockForZeroCmp, const DataLayout &TheDataLayout);
110 
111  unsigned getNumBlocks();
112  uint64_t getNumLoads() const { return LoadSequence.size(); }
113 
114  Value *getMemCmpExpansion();
115 };
116 
117 MemCmpExpansion::LoadEntryVector MemCmpExpansion::computeGreedyLoadSequence(
118  uint64_t Size, llvm::ArrayRef<unsigned> LoadSizes,
119  const unsigned MaxNumLoads, unsigned &NumLoadsNonOneByte) {
120  NumLoadsNonOneByte = 0;
121  LoadEntryVector LoadSequence;
122  uint64_t Offset = 0;
123  while (Size && !LoadSizes.empty()) {
124  const unsigned LoadSize = LoadSizes.front();
125  const uint64_t NumLoadsForThisSize = Size / LoadSize;
126  if (LoadSequence.size() + NumLoadsForThisSize > MaxNumLoads) {
127  // Do not expand if the total number of loads is larger than what the
128  // target allows. Note that it's important that we exit before completing
129  // the expansion to avoid using a ton of memory to store the expansion for
130  // large sizes.
131  return {};
132  }
133  if (NumLoadsForThisSize > 0) {
134  for (uint64_t I = 0; I < NumLoadsForThisSize; ++I) {
135  LoadSequence.push_back({LoadSize, Offset});
136  Offset += LoadSize;
137  }
138  if (LoadSize > 1)
139  ++NumLoadsNonOneByte;
140  Size = Size % LoadSize;
141  }
142  LoadSizes = LoadSizes.drop_front();
143  }
144  return LoadSequence;
145 }
146 
148 MemCmpExpansion::computeOverlappingLoadSequence(uint64_t Size,
149  const unsigned MaxLoadSize,
150  const unsigned MaxNumLoads,
151  unsigned &NumLoadsNonOneByte) {
152  // These are already handled by the greedy approach.
153  if (Size < 2 || MaxLoadSize < 2)
154  return {};
155 
156  // We try to do as many non-overlapping loads as possible starting from the
157  // beginning.
158  const uint64_t NumNonOverlappingLoads = Size / MaxLoadSize;
159  assert(NumNonOverlappingLoads && "there must be at least one load");
160  // There remain 0 to (MaxLoadSize - 1) bytes to load, this will be done with
161  // an overlapping load.
162  Size = Size - NumNonOverlappingLoads * MaxLoadSize;
163  // Bail if we do not need an overloapping store, this is already handled by
164  // the greedy approach.
165  if (Size == 0)
166  return {};
167  // Bail if the number of loads (non-overlapping + potential overlapping one)
168  // is larger than the max allowed.
169  if ((NumNonOverlappingLoads + 1) > MaxNumLoads)
170  return {};
171 
172  // Add non-overlapping loads.
173  LoadEntryVector LoadSequence;
174  uint64_t Offset = 0;
175  for (uint64_t I = 0; I < NumNonOverlappingLoads; ++I) {
176  LoadSequence.push_back({MaxLoadSize, Offset});
177  Offset += MaxLoadSize;
178  }
179 
180  // Add the last overlapping load.
181  assert(Size > 0 && Size < MaxLoadSize && "broken invariant");
182  LoadSequence.push_back({MaxLoadSize, Offset - (MaxLoadSize - Size)});
183  NumLoadsNonOneByte = 1;
184  return LoadSequence;
185 }
186 
187 // Initialize the basic block structure required for expansion of memcmp call
188 // with given maximum load size and memcmp size parameter.
189 // This structure includes:
190 // 1. A list of load compare blocks - LoadCmpBlocks.
191 // 2. An EndBlock, split from original instruction point, which is the block to
192 // return from.
193 // 3. ResultBlock, block to branch to for early exit when a
194 // LoadCmpBlock finds a difference.
195 MemCmpExpansion::MemCmpExpansion(
196  CallInst *const CI, uint64_t Size,
198  const unsigned MaxNumLoads, const bool IsUsedForZeroCmp,
199  const unsigned MaxLoadsPerBlockForZeroCmp, const DataLayout &TheDataLayout)
200  : CI(CI),
201  Size(Size),
202  MaxLoadSize(0),
203  NumLoadsNonOneByte(0),
204  NumLoadsPerBlockForZeroCmp(MaxLoadsPerBlockForZeroCmp),
205  IsUsedForZeroCmp(IsUsedForZeroCmp),
206  DL(TheDataLayout),
207  Builder(CI) {
208  assert(Size > 0 && "zero blocks");
209  // Scale the max size down if the target can load more bytes than we need.
210  llvm::ArrayRef<unsigned> LoadSizes(Options.LoadSizes);
211  while (!LoadSizes.empty() && LoadSizes.front() > Size) {
212  LoadSizes = LoadSizes.drop_front();
213  }
214  assert(!LoadSizes.empty() && "cannot load Size bytes");
215  MaxLoadSize = LoadSizes.front();
216  // Compute the decomposition.
217  unsigned GreedyNumLoadsNonOneByte = 0;
218  LoadSequence = computeGreedyLoadSequence(Size, LoadSizes, MaxNumLoads,
219  GreedyNumLoadsNonOneByte);
220  NumLoadsNonOneByte = GreedyNumLoadsNonOneByte;
221  assert(LoadSequence.size() <= MaxNumLoads && "broken invariant");
222  // If we allow overlapping loads and the load sequence is not already optimal,
223  // use overlapping loads.
224  if (Options.AllowOverlappingLoads &&
225  (LoadSequence.empty() || LoadSequence.size() > 2)) {
226  unsigned OverlappingNumLoadsNonOneByte = 0;
227  auto OverlappingLoads = computeOverlappingLoadSequence(
228  Size, MaxLoadSize, MaxNumLoads, OverlappingNumLoadsNonOneByte);
229  if (!OverlappingLoads.empty() &&
230  (LoadSequence.empty() ||
231  OverlappingLoads.size() < LoadSequence.size())) {
232  LoadSequence = OverlappingLoads;
233  NumLoadsNonOneByte = OverlappingNumLoadsNonOneByte;
234  }
235  }
236  assert(LoadSequence.size() <= MaxNumLoads && "broken invariant");
237 }
238 
239 unsigned MemCmpExpansion::getNumBlocks() {
240  if (IsUsedForZeroCmp)
241  return getNumLoads() / NumLoadsPerBlockForZeroCmp +
242  (getNumLoads() % NumLoadsPerBlockForZeroCmp != 0 ? 1 : 0);
243  return getNumLoads();
244 }
245 
246 void MemCmpExpansion::createLoadCmpBlocks() {
247  for (unsigned i = 0; i < getNumBlocks(); i++) {
248  BasicBlock *BB = BasicBlock::Create(CI->getContext(), "loadbb",
249  EndBlock->getParent(), EndBlock);
250  LoadCmpBlocks.push_back(BB);
251  }
252 }
253 
254 void MemCmpExpansion::createResultBlock() {
255  ResBlock.BB = BasicBlock::Create(CI->getContext(), "res_block",
256  EndBlock->getParent(), EndBlock);
257 }
258 
259 /// Return a pointer to an element of type `LoadSizeType` at offset
260 /// `OffsetBytes`.
261 Value *MemCmpExpansion::getPtrToElementAtOffset(Value *Source,
262  Type *LoadSizeType,
263  uint64_t OffsetBytes) {
264  if (OffsetBytes > 0) {
265  auto *ByteType = Type::getInt8Ty(CI->getContext());
266  Source = Builder.CreateGEP(
267  ByteType, Builder.CreateBitCast(Source, ByteType->getPointerTo()),
268  ConstantInt::get(ByteType, OffsetBytes));
269  }
270  return Builder.CreateBitCast(Source, LoadSizeType->getPointerTo());
271 }
272 
273 // This function creates the IR instructions for loading and comparing 1 byte.
274 // It loads 1 byte from each source of the memcmp parameters with the given
275 // GEPIndex. It then subtracts the two loaded values and adds this result to the
276 // final phi node for selecting the memcmp result.
277 void MemCmpExpansion::emitLoadCompareByteBlock(unsigned BlockIndex,
278  unsigned OffsetBytes) {
279  Builder.SetInsertPoint(LoadCmpBlocks[BlockIndex]);
280  Type *LoadSizeType = Type::getInt8Ty(CI->getContext());
281  Value *Source1 =
282  getPtrToElementAtOffset(CI->getArgOperand(0), LoadSizeType, OffsetBytes);
283  Value *Source2 =
284  getPtrToElementAtOffset(CI->getArgOperand(1), LoadSizeType, OffsetBytes);
285 
286  Value *LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
287  Value *LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
288 
289  LoadSrc1 = Builder.CreateZExt(LoadSrc1, Type::getInt32Ty(CI->getContext()));
290  LoadSrc2 = Builder.CreateZExt(LoadSrc2, Type::getInt32Ty(CI->getContext()));
291  Value *Diff = Builder.CreateSub(LoadSrc1, LoadSrc2);
292 
293  PhiRes->addIncoming(Diff, LoadCmpBlocks[BlockIndex]);
294 
295  if (BlockIndex < (LoadCmpBlocks.size() - 1)) {
296  // Early exit branch if difference found to EndBlock. Otherwise, continue to
297  // next LoadCmpBlock,
298  Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_NE, Diff,
299  ConstantInt::get(Diff->getType(), 0));
300  BranchInst *CmpBr =
301  BranchInst::Create(EndBlock, LoadCmpBlocks[BlockIndex + 1], Cmp);
302  Builder.Insert(CmpBr);
303  } else {
304  // The last block has an unconditional branch to EndBlock.
305  BranchInst *CmpBr = BranchInst::Create(EndBlock);
306  Builder.Insert(CmpBr);
307  }
308 }
309 
310 /// Generate an equality comparison for one or more pairs of loaded values.
311 /// This is used in the case where the memcmp() call is compared equal or not
312 /// equal to zero.
313 Value *MemCmpExpansion::getCompareLoadPairs(unsigned BlockIndex,
314  unsigned &LoadIndex) {
315  assert(LoadIndex < getNumLoads() &&
316  "getCompareLoadPairs() called with no remaining loads");
317  std::vector<Value *> XorList, OrList;
318  Value *Diff;
319 
320  const unsigned NumLoads =
321  std::min(getNumLoads() - LoadIndex, NumLoadsPerBlockForZeroCmp);
322 
323  // For a single-block expansion, start inserting before the memcmp call.
324  if (LoadCmpBlocks.empty())
325  Builder.SetInsertPoint(CI);
326  else
327  Builder.SetInsertPoint(LoadCmpBlocks[BlockIndex]);
328 
329  Value *Cmp = nullptr;
330  // If we have multiple loads per block, we need to generate a composite
331  // comparison using xor+or. The type for the combinations is the largest load
332  // type.
333  IntegerType *const MaxLoadType =
334  NumLoads == 1 ? nullptr
335  : IntegerType::get(CI->getContext(), MaxLoadSize * 8);
336  for (unsigned i = 0; i < NumLoads; ++i, ++LoadIndex) {
337  const LoadEntry &CurLoadEntry = LoadSequence[LoadIndex];
338 
339  IntegerType *LoadSizeType =
340  IntegerType::get(CI->getContext(), CurLoadEntry.LoadSize * 8);
341 
342  Value *Source1 = getPtrToElementAtOffset(CI->getArgOperand(0), LoadSizeType,
343  CurLoadEntry.Offset);
344  Value *Source2 = getPtrToElementAtOffset(CI->getArgOperand(1), LoadSizeType,
345  CurLoadEntry.Offset);
346 
347  // Get a constant or load a value for each source address.
348  Value *LoadSrc1 = nullptr;
349  if (auto *Source1C = dyn_cast<Constant>(Source1))
350  LoadSrc1 = ConstantFoldLoadFromConstPtr(Source1C, LoadSizeType, DL);
351  if (!LoadSrc1)
352  LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
353 
354  Value *LoadSrc2 = nullptr;
355  if (auto *Source2C = dyn_cast<Constant>(Source2))
356  LoadSrc2 = ConstantFoldLoadFromConstPtr(Source2C, LoadSizeType, DL);
357  if (!LoadSrc2)
358  LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
359 
360  if (NumLoads != 1) {
361  if (LoadSizeType != MaxLoadType) {
362  LoadSrc1 = Builder.CreateZExt(LoadSrc1, MaxLoadType);
363  LoadSrc2 = Builder.CreateZExt(LoadSrc2, MaxLoadType);
364  }
365  // If we have multiple loads per block, we need to generate a composite
366  // comparison using xor+or.
367  Diff = Builder.CreateXor(LoadSrc1, LoadSrc2);
368  Diff = Builder.CreateZExt(Diff, MaxLoadType);
369  XorList.push_back(Diff);
370  } else {
371  // If there's only one load per block, we just compare the loaded values.
372  Cmp = Builder.CreateICmpNE(LoadSrc1, LoadSrc2);
373  }
374  }
375 
376  auto pairWiseOr = [&](std::vector<Value *> &InList) -> std::vector<Value *> {
377  std::vector<Value *> OutList;
378  for (unsigned i = 0; i < InList.size() - 1; i = i + 2) {
379  Value *Or = Builder.CreateOr(InList[i], InList[i + 1]);
380  OutList.push_back(Or);
381  }
382  if (InList.size() % 2 != 0)
383  OutList.push_back(InList.back());
384  return OutList;
385  };
386 
387  if (!Cmp) {
388  // Pairwise OR the XOR results.
389  OrList = pairWiseOr(XorList);
390 
391  // Pairwise OR the OR results until one result left.
392  while (OrList.size() != 1) {
393  OrList = pairWiseOr(OrList);
394  }
395  Cmp = Builder.CreateICmpNE(OrList[0], ConstantInt::get(Diff->getType(), 0));
396  }
397 
398  return Cmp;
399 }
400 
401 void MemCmpExpansion::emitLoadCompareBlockMultipleLoads(unsigned BlockIndex,
402  unsigned &LoadIndex) {
403  Value *Cmp = getCompareLoadPairs(BlockIndex, LoadIndex);
404 
405  BasicBlock *NextBB = (BlockIndex == (LoadCmpBlocks.size() - 1))
406  ? EndBlock
407  : LoadCmpBlocks[BlockIndex + 1];
408  // Early exit branch if difference found to ResultBlock. Otherwise,
409  // continue to next LoadCmpBlock or EndBlock.
410  BranchInst *CmpBr = BranchInst::Create(ResBlock.BB, NextBB, Cmp);
411  Builder.Insert(CmpBr);
412 
413  // Add a phi edge for the last LoadCmpBlock to Endblock with a value of 0
414  // since early exit to ResultBlock was not taken (no difference was found in
415  // any of the bytes).
416  if (BlockIndex == LoadCmpBlocks.size() - 1) {
417  Value *Zero = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 0);
418  PhiRes->addIncoming(Zero, LoadCmpBlocks[BlockIndex]);
419  }
420 }
421 
422 // This function creates the IR intructions for loading and comparing using the
423 // given LoadSize. It loads the number of bytes specified by LoadSize from each
424 // source of the memcmp parameters. It then does a subtract to see if there was
425 // a difference in the loaded values. If a difference is found, it branches
426 // with an early exit to the ResultBlock for calculating which source was
427 // larger. Otherwise, it falls through to the either the next LoadCmpBlock or
428 // the EndBlock if this is the last LoadCmpBlock. Loading 1 byte is handled with
429 // a special case through emitLoadCompareByteBlock. The special handling can
430 // simply subtract the loaded values and add it to the result phi node.
431 void MemCmpExpansion::emitLoadCompareBlock(unsigned BlockIndex) {
432  // There is one load per block in this case, BlockIndex == LoadIndex.
433  const LoadEntry &CurLoadEntry = LoadSequence[BlockIndex];
434 
435  if (CurLoadEntry.LoadSize == 1) {
436  MemCmpExpansion::emitLoadCompareByteBlock(BlockIndex, CurLoadEntry.Offset);
437  return;
438  }
439 
440  Type *LoadSizeType =
441  IntegerType::get(CI->getContext(), CurLoadEntry.LoadSize * 8);
442  Type *MaxLoadType = IntegerType::get(CI->getContext(), MaxLoadSize * 8);
443  assert(CurLoadEntry.LoadSize <= MaxLoadSize && "Unexpected load type");
444 
445  Builder.SetInsertPoint(LoadCmpBlocks[BlockIndex]);
446 
447  Value *Source1 = getPtrToElementAtOffset(CI->getArgOperand(0), LoadSizeType,
448  CurLoadEntry.Offset);
449  Value *Source2 = getPtrToElementAtOffset(CI->getArgOperand(1), LoadSizeType,
450  CurLoadEntry.Offset);
451 
452  // Load LoadSizeType from the base address.
453  Value *LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
454  Value *LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
455 
456  if (DL.isLittleEndian()) {
458  Intrinsic::bswap, LoadSizeType);
459  LoadSrc1 = Builder.CreateCall(Bswap, LoadSrc1);
460  LoadSrc2 = Builder.CreateCall(Bswap, LoadSrc2);
461  }
462 
463  if (LoadSizeType != MaxLoadType) {
464  LoadSrc1 = Builder.CreateZExt(LoadSrc1, MaxLoadType);
465  LoadSrc2 = Builder.CreateZExt(LoadSrc2, MaxLoadType);
466  }
467 
468  // Add the loaded values to the phi nodes for calculating memcmp result only
469  // if result is not used in a zero equality.
470  if (!IsUsedForZeroCmp) {
471  ResBlock.PhiSrc1->addIncoming(LoadSrc1, LoadCmpBlocks[BlockIndex]);
472  ResBlock.PhiSrc2->addIncoming(LoadSrc2, LoadCmpBlocks[BlockIndex]);
473  }
474 
475  Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, LoadSrc1, LoadSrc2);
476  BasicBlock *NextBB = (BlockIndex == (LoadCmpBlocks.size() - 1))
477  ? EndBlock
478  : LoadCmpBlocks[BlockIndex + 1];
479  // Early exit branch if difference found to ResultBlock. Otherwise, continue
480  // to next LoadCmpBlock or EndBlock.
481  BranchInst *CmpBr = BranchInst::Create(NextBB, ResBlock.BB, Cmp);
482  Builder.Insert(CmpBr);
483 
484  // Add a phi edge for the last LoadCmpBlock to Endblock with a value of 0
485  // since early exit to ResultBlock was not taken (no difference was found in
486  // any of the bytes).
487  if (BlockIndex == LoadCmpBlocks.size() - 1) {
488  Value *Zero = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 0);
489  PhiRes->addIncoming(Zero, LoadCmpBlocks[BlockIndex]);
490  }
491 }
492 
493 // This function populates the ResultBlock with a sequence to calculate the
494 // memcmp result. It compares the two loaded source values and returns -1 if
495 // src1 < src2 and 1 if src1 > src2.
496 void MemCmpExpansion::emitMemCmpResultBlock() {
497  // Special case: if memcmp result is used in a zero equality, result does not
498  // need to be calculated and can simply return 1.
499  if (IsUsedForZeroCmp) {
500  BasicBlock::iterator InsertPt = ResBlock.BB->getFirstInsertionPt();
501  Builder.SetInsertPoint(ResBlock.BB, InsertPt);
502  Value *Res = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 1);
503  PhiRes->addIncoming(Res, ResBlock.BB);
504  BranchInst *NewBr = BranchInst::Create(EndBlock);
505  Builder.Insert(NewBr);
506  return;
507  }
508  BasicBlock::iterator InsertPt = ResBlock.BB->getFirstInsertionPt();
509  Builder.SetInsertPoint(ResBlock.BB, InsertPt);
510 
511  Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_ULT, ResBlock.PhiSrc1,
512  ResBlock.PhiSrc2);
513 
514  Value *Res =
515  Builder.CreateSelect(Cmp, ConstantInt::get(Builder.getInt32Ty(), -1),
516  ConstantInt::get(Builder.getInt32Ty(), 1));
517 
518  BranchInst *NewBr = BranchInst::Create(EndBlock);
519  Builder.Insert(NewBr);
520  PhiRes->addIncoming(Res, ResBlock.BB);
521 }
522 
523 void MemCmpExpansion::setupResultBlockPHINodes() {
524  Type *MaxLoadType = IntegerType::get(CI->getContext(), MaxLoadSize * 8);
525  Builder.SetInsertPoint(ResBlock.BB);
526  // Note: this assumes one load per block.
527  ResBlock.PhiSrc1 =
528  Builder.CreatePHI(MaxLoadType, NumLoadsNonOneByte, "phi.src1");
529  ResBlock.PhiSrc2 =
530  Builder.CreatePHI(MaxLoadType, NumLoadsNonOneByte, "phi.src2");
531 }
532 
533 void MemCmpExpansion::setupEndBlockPHINodes() {
534  Builder.SetInsertPoint(&EndBlock->front());
535  PhiRes = Builder.CreatePHI(Type::getInt32Ty(CI->getContext()), 2, "phi.res");
536 }
537 
538 Value *MemCmpExpansion::getMemCmpExpansionZeroCase() {
539  unsigned LoadIndex = 0;
540  // This loop populates each of the LoadCmpBlocks with the IR sequence to
541  // handle multiple loads per block.
542  for (unsigned I = 0; I < getNumBlocks(); ++I) {
543  emitLoadCompareBlockMultipleLoads(I, LoadIndex);
544  }
545 
546  emitMemCmpResultBlock();
547  return PhiRes;
548 }
549 
550 /// A memcmp expansion that compares equality with 0 and only has one block of
551 /// load and compare can bypass the compare, branch, and phi IR that is required
552 /// in the general case.
553 Value *MemCmpExpansion::getMemCmpEqZeroOneBlock() {
554  unsigned LoadIndex = 0;
555  Value *Cmp = getCompareLoadPairs(0, LoadIndex);
556  assert(LoadIndex == getNumLoads() && "some entries were not consumed");
557  return Builder.CreateZExt(Cmp, Type::getInt32Ty(CI->getContext()));
558 }
559 
560 /// A memcmp expansion that only has one block of load and compare can bypass
561 /// the compare, branch, and phi IR that is required in the general case.
562 Value *MemCmpExpansion::getMemCmpOneBlock() {
563  Type *LoadSizeType = IntegerType::get(CI->getContext(), Size * 8);
564  Value *Source1 = CI->getArgOperand(0);
565  Value *Source2 = CI->getArgOperand(1);
566 
567  // Cast source to LoadSizeType*.
568  if (Source1->getType() != LoadSizeType)
569  Source1 = Builder.CreateBitCast(Source1, LoadSizeType->getPointerTo());
570  if (Source2->getType() != LoadSizeType)
571  Source2 = Builder.CreateBitCast(Source2, LoadSizeType->getPointerTo());
572 
573  // Load LoadSizeType from the base address.
574  Value *LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
575  Value *LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
576 
577  if (DL.isLittleEndian() && Size != 1) {
579  Intrinsic::bswap, LoadSizeType);
580  LoadSrc1 = Builder.CreateCall(Bswap, LoadSrc1);
581  LoadSrc2 = Builder.CreateCall(Bswap, LoadSrc2);
582  }
583 
584  if (Size < 4) {
585  // The i8 and i16 cases don't need compares. We zext the loaded values and
586  // subtract them to get the suitable negative, zero, or positive i32 result.
587  LoadSrc1 = Builder.CreateZExt(LoadSrc1, Builder.getInt32Ty());
588  LoadSrc2 = Builder.CreateZExt(LoadSrc2, Builder.getInt32Ty());
589  return Builder.CreateSub(LoadSrc1, LoadSrc2);
590  }
591 
592  // The result of memcmp is negative, zero, or positive, so produce that by
593  // subtracting 2 extended compare bits: sub (ugt, ult).
594  // If a target prefers to use selects to get -1/0/1, they should be able
595  // to transform this later. The inverse transform (going from selects to math)
596  // may not be possible in the DAG because the selects got converted into
597  // branches before we got there.
598  Value *CmpUGT = Builder.CreateICmpUGT(LoadSrc1, LoadSrc2);
599  Value *CmpULT = Builder.CreateICmpULT(LoadSrc1, LoadSrc2);
600  Value *ZextUGT = Builder.CreateZExt(CmpUGT, Builder.getInt32Ty());
601  Value *ZextULT = Builder.CreateZExt(CmpULT, Builder.getInt32Ty());
602  return Builder.CreateSub(ZextUGT, ZextULT);
603 }
604 
605 // This function expands the memcmp call into an inline expansion and returns
606 // the memcmp result.
607 Value *MemCmpExpansion::getMemCmpExpansion() {
608  // Create the basic block framework for a multi-block expansion.
609  if (getNumBlocks() != 1) {
610  BasicBlock *StartBlock = CI->getParent();
611  EndBlock = StartBlock->splitBasicBlock(CI, "endblock");
612  setupEndBlockPHINodes();
613  createResultBlock();
614 
615  // If return value of memcmp is not used in a zero equality, we need to
616  // calculate which source was larger. The calculation requires the
617  // two loaded source values of each load compare block.
618  // These will be saved in the phi nodes created by setupResultBlockPHINodes.
619  if (!IsUsedForZeroCmp) setupResultBlockPHINodes();
620 
621  // Create the number of required load compare basic blocks.
622  createLoadCmpBlocks();
623 
624  // Update the terminator added by splitBasicBlock to branch to the first
625  // LoadCmpBlock.
626  StartBlock->getTerminator()->setSuccessor(0, LoadCmpBlocks[0]);
627  }
628 
629  Builder.SetCurrentDebugLocation(CI->getDebugLoc());
630 
631  if (IsUsedForZeroCmp)
632  return getNumBlocks() == 1 ? getMemCmpEqZeroOneBlock()
633  : getMemCmpExpansionZeroCase();
634 
635  if (getNumBlocks() == 1)
636  return getMemCmpOneBlock();
637 
638  for (unsigned I = 0; I < getNumBlocks(); ++I) {
639  emitLoadCompareBlock(I);
640  }
641 
642  emitMemCmpResultBlock();
643  return PhiRes;
644 }
645 
646 // This function checks to see if an expansion of memcmp can be generated.
647 // It checks for constant compare size that is less than the max inline size.
648 // If an expansion cannot occur, returns false to leave as a library call.
649 // Otherwise, the library call is replaced with a new IR instruction sequence.
650 /// We want to transform:
651 /// %call = call signext i32 @memcmp(i8* %0, i8* %1, i64 15)
652 /// To:
653 /// loadbb:
654 /// %0 = bitcast i32* %buffer2 to i8*
655 /// %1 = bitcast i32* %buffer1 to i8*
656 /// %2 = bitcast i8* %1 to i64*
657 /// %3 = bitcast i8* %0 to i64*
658 /// %4 = load i64, i64* %2
659 /// %5 = load i64, i64* %3
660 /// %6 = call i64 @llvm.bswap.i64(i64 %4)
661 /// %7 = call i64 @llvm.bswap.i64(i64 %5)
662 /// %8 = sub i64 %6, %7
663 /// %9 = icmp ne i64 %8, 0
664 /// br i1 %9, label %res_block, label %loadbb1
665 /// res_block: ; preds = %loadbb2,
666 /// %loadbb1, %loadbb
667 /// %phi.src1 = phi i64 [ %6, %loadbb ], [ %22, %loadbb1 ], [ %36, %loadbb2 ]
668 /// %phi.src2 = phi i64 [ %7, %loadbb ], [ %23, %loadbb1 ], [ %37, %loadbb2 ]
669 /// %10 = icmp ult i64 %phi.src1, %phi.src2
670 /// %11 = select i1 %10, i32 -1, i32 1
671 /// br label %endblock
672 /// loadbb1: ; preds = %loadbb
673 /// %12 = bitcast i32* %buffer2 to i8*
674 /// %13 = bitcast i32* %buffer1 to i8*
675 /// %14 = bitcast i8* %13 to i32*
676 /// %15 = bitcast i8* %12 to i32*
677 /// %16 = getelementptr i32, i32* %14, i32 2
678 /// %17 = getelementptr i32, i32* %15, i32 2
679 /// %18 = load i32, i32* %16
680 /// %19 = load i32, i32* %17
681 /// %20 = call i32 @llvm.bswap.i32(i32 %18)
682 /// %21 = call i32 @llvm.bswap.i32(i32 %19)
683 /// %22 = zext i32 %20 to i64
684 /// %23 = zext i32 %21 to i64
685 /// %24 = sub i64 %22, %23
686 /// %25 = icmp ne i64 %24, 0
687 /// br i1 %25, label %res_block, label %loadbb2
688 /// loadbb2: ; preds = %loadbb1
689 /// %26 = bitcast i32* %buffer2 to i8*
690 /// %27 = bitcast i32* %buffer1 to i8*
691 /// %28 = bitcast i8* %27 to i16*
692 /// %29 = bitcast i8* %26 to i16*
693 /// %30 = getelementptr i16, i16* %28, i16 6
694 /// %31 = getelementptr i16, i16* %29, i16 6
695 /// %32 = load i16, i16* %30
696 /// %33 = load i16, i16* %31
697 /// %34 = call i16 @llvm.bswap.i16(i16 %32)
698 /// %35 = call i16 @llvm.bswap.i16(i16 %33)
699 /// %36 = zext i16 %34 to i64
700 /// %37 = zext i16 %35 to i64
701 /// %38 = sub i64 %36, %37
702 /// %39 = icmp ne i64 %38, 0
703 /// br i1 %39, label %res_block, label %loadbb3
704 /// loadbb3: ; preds = %loadbb2
705 /// %40 = bitcast i32* %buffer2 to i8*
706 /// %41 = bitcast i32* %buffer1 to i8*
707 /// %42 = getelementptr i8, i8* %41, i8 14
708 /// %43 = getelementptr i8, i8* %40, i8 14
709 /// %44 = load i8, i8* %42
710 /// %45 = load i8, i8* %43
711 /// %46 = zext i8 %44 to i32
712 /// %47 = zext i8 %45 to i32
713 /// %48 = sub i32 %46, %47
714 /// br label %endblock
715 /// endblock: ; preds = %res_block,
716 /// %loadbb3
717 /// %phi.res = phi i32 [ %48, %loadbb3 ], [ %11, %res_block ]
718 /// ret i32 %phi.res
719 static bool expandMemCmp(CallInst *CI, const TargetTransformInfo *TTI,
720  const TargetLowering *TLI, const DataLayout *DL) {
721  NumMemCmpCalls++;
722 
723  // Early exit from expansion if -Oz.
724  if (CI->getFunction()->optForMinSize())
725  return false;
726 
727  // Early exit from expansion if size is not a constant.
728  ConstantInt *SizeCast = dyn_cast<ConstantInt>(CI->getArgOperand(2));
729  if (!SizeCast) {
730  NumMemCmpNotConstant++;
731  return false;
732  }
733  const uint64_t SizeVal = SizeCast->getZExtValue();
734 
735  if (SizeVal == 0) {
736  return false;
737  }
738  // TTI call to check if target would like to expand memcmp. Also, get the
739  // available load sizes.
740  const bool IsUsedForZeroCmp = isOnlyUsedInZeroEqualityComparison(CI);
741  const auto *const Options = TTI->enableMemCmpExpansion(IsUsedForZeroCmp);
742  if (!Options) return false;
743 
744  const unsigned MaxNumLoads =
746 
747  unsigned NumLoadsPerBlock = MemCmpEqZeroNumLoadsPerBlock.getNumOccurrences()
750 
751  MemCmpExpansion Expansion(CI, SizeVal, *Options, MaxNumLoads,
752  IsUsedForZeroCmp, NumLoadsPerBlock, *DL);
753 
754  // Don't expand if this will require more loads than desired by the target.
755  if (Expansion.getNumLoads() == 0) {
756  NumMemCmpGreaterThanMax++;
757  return false;
758  }
759 
760  NumMemCmpInlined++;
761 
762  Value *Res = Expansion.getMemCmpExpansion();
763 
764  // Replace call with result of expansion and erase call.
765  CI->replaceAllUsesWith(Res);
766  CI->eraseFromParent();
767 
768  return true;
769 }
770 
771 
772 
773 class ExpandMemCmpPass : public FunctionPass {
774 public:
775  static char ID;
776 
777  ExpandMemCmpPass() : FunctionPass(ID) {
778  initializeExpandMemCmpPassPass(*PassRegistry::getPassRegistry());
779  }
780 
781  bool runOnFunction(Function &F) override {
782  if (skipFunction(F)) return false;
783 
784  auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
785  if (!TPC) {
786  return false;
787  }
788  const TargetLowering* TL =
789  TPC->getTM<TargetMachine>().getSubtargetImpl(F)->getTargetLowering();
790 
791  const TargetLibraryInfo *TLI =
792  &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
793  const TargetTransformInfo *TTI =
794  &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
795  auto PA = runImpl(F, TLI, TTI, TL);
796  return !PA.areAllPreserved();
797  }
798 
799 private:
800  void getAnalysisUsage(AnalysisUsage &AU) const override {
804  }
805 
807  const TargetTransformInfo *TTI,
808  const TargetLowering* TL);
809  // Returns true if a change was made.
810  bool runOnBlock(BasicBlock &BB, const TargetLibraryInfo *TLI,
811  const TargetTransformInfo *TTI, const TargetLowering* TL,
812  const DataLayout& DL);
813 };
814 
815 bool ExpandMemCmpPass::runOnBlock(
816  BasicBlock &BB, const TargetLibraryInfo *TLI,
817  const TargetTransformInfo *TTI, const TargetLowering* TL,
818  const DataLayout& DL) {
819  for (Instruction& I : BB) {
820  CallInst *CI = dyn_cast<CallInst>(&I);
821  if (!CI) {
822  continue;
823  }
824  LibFunc Func;
825  if (TLI->getLibFunc(ImmutableCallSite(CI), Func) &&
826  Func == LibFunc_memcmp && expandMemCmp(CI, TTI, TL, &DL)) {
827  return true;
828  }
829  }
830  return false;
831 }
832 
833 
835  Function &F, const TargetLibraryInfo *TLI, const TargetTransformInfo *TTI,
836  const TargetLowering* TL) {
837  const DataLayout& DL = F.getParent()->getDataLayout();
838  bool MadeChanges = false;
839  for (auto BBIt = F.begin(); BBIt != F.end();) {
840  if (runOnBlock(*BBIt, TLI, TTI, TL, DL)) {
841  MadeChanges = true;
842  // If changes were made, restart the function from the beginning, since
843  // the structure of the function was changed.
844  BBIt = F.begin();
845  } else {
846  ++BBIt;
847  }
848  }
849  return MadeChanges ? PreservedAnalyses::none() : PreservedAnalyses::all();
850 }
851 
852 } // namespace
853 
854 char ExpandMemCmpPass::ID = 0;
855 INITIALIZE_PASS_BEGIN(ExpandMemCmpPass, "expandmemcmp",
856  "Expand memcmp() to load/stores", false, false)
859 INITIALIZE_PASS_END(ExpandMemCmpPass, "expandmemcmp",
860  "Expand memcmp() to load/stores", false, false)
861 
863  return new ExpandMemCmpPass();
864 }
const T & front() const
front - Get the first element.
Definition: ArrayRef.h:151
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:67
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:110
Value * CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1984
static bool runImpl(Function &F, TargetLibraryInfo &TLI, DominatorTree &DT)
This is the entry point for all transforms.
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
bool isOnlyUsedInZeroEqualityComparison(const Instruction *CxtI)
Value * CreateICmpNE(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1878
This class represents lattice values for constants.
Definition: AllocatorList.h:23
Value * CreateICmpULT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1890
LoadInst * CreateLoad(Type *Ty, Value *Ptr, const char *Name)
Provided to resolve &#39;CreateLoad(Ty, Ptr, "...")&#39; correctly, instead of converting the string to &#39;bool...
Definition: IRBuilder.h:1392
Value * CreateXor(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1235
iterator end()
Definition: Function.h:657
INITIALIZE_PASS_BEGIN(ExpandMemCmpPass, "expandmemcmp", "Expand memcmp() to load/stores", false, false) INITIALIZE_PASS_END(ExpandMemCmpPass
This class represents a function call, abstracting a target machine&#39;s calling convention.
FunctionPass * createExpandMemCmpPass()
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:705
STATISTIC(NumFunctions, "Total number of functions")
F(f)
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:137
void setSuccessor(unsigned Idx, BasicBlock *BB)
Update the specified successor to point at the provided block.
IntegerType * getInt32Ty()
Fetch the type representing a 32-bit integer.
Definition: IRBuilder.h:346
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1155
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:50
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:369
PointerType * getPointerTo(unsigned AddrSpace=0) const
Return a pointer to the current type.
Definition: Type.cpp:651
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:742
void initializeExpandMemCmpPassPass(PassRegistry &)
This class defines information used to lower LLVM code to legal SelectionDAG operators that the targe...
Value * CreateBitCast(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1767
virtual void getAnalysisUsage(AnalysisUsage &) const
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: Pass.cpp:91
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:244
expandmemcmp
Value * CreateICmpUGT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1882
Predicate all(Predicate P0, Predicate P1)
True iff P0 and P1 are true.
bool isLittleEndian() const
Layout endianness...
Definition: DataLayout.h:220
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1066
void SetCurrentDebugLocation(DebugLoc L)
Set location information used by debugging information.
Definition: IRBuilder.h:150
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:429
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1694
iterator begin()
Definition: Function.h:655
Function * getDeclaration(Module *M, ID id, ArrayRef< Type *> Tys=None)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:1018
void SetInsertPoint(BasicBlock *TheBB)
This specifies that created instructions should be appended to the end of the specified block...
Definition: IRBuilder.h:126
static PreservedAnalyses none()
Convenience factory function for the empty preserved set.
Definition: PassManager.h:156
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1217
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:422
If not nullptr, enable inline expansion of memcmp.
uint64_t getZExtValue() const
Return the constant as a 64-bit unsigned integer value after it has been zero extended as appropriate...
Definition: Constants.h:148
Wrapper pass for TargetTransformInfo.
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:45
Conditional or Unconditional Branch instruction.
Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)
Definition: IRBuilder.h:2057
const Instruction & front() const
Definition: BasicBlock.h:280
Represent the analysis usage information of a pass.
bool optForSize() const
Optimize this function for size (-Os) or minimum size (-Oz).
Definition: Function.h:597
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:284
Class to represent integer types.
Definition: DerivedTypes.h:39
Expand memcmp() to load/stores"
const Function * getFunction() const
Return the function this instruction belongs to.
Definition: Instruction.cpp:59
Constant * ConstantFoldLoadFromConstPtr(Constant *C, Type *Ty, const DataLayout &DL)
ConstantFoldLoadFromConstPtr - Return the value that a load from C would produce if it is constant an...
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
PHINode * CreatePHI(Type *Ty, unsigned NumReservedValues, const Twine &Name="")
Definition: IRBuilder.h:2004
Value * CreateGEP(Value *Ptr, ArrayRef< Value *> IdxList, const Twine &Name="")
Definition: IRBuilder.h:1493
Iterator for intrusive lists based on ilist_node.
This is the shared class of boolean and integer constants.
Definition: Constants.h:83
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
static cl::opt< unsigned > MemCmpEqZeroNumLoadsPerBlock("memcmp-num-loads-per-block", cl::Hidden, cl::init(1), cl::desc("The number of loads per basic block for inline expansion of " "memcmp that is only being compared against zero."))
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:839
Provides information about what library functions are available for the current target.
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:55
unsigned getMaxExpandSizeMemcmp(bool OptSize) const
Get maximum # of load operations permitted for memcmp.
bool getLibFunc(StringRef funcName, LibFunc &F) const
Searches for a particular function name.
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:324
ArrayRef< T > drop_front(size_t N=1) const
Drop the first N elements of the array.
Definition: ArrayRef.h:187
Establish a view to a call site for examination.
Definition: CallSite.h:892
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:106
#define I(x, y, z)
Definition: MD5.cpp:58
bool optForMinSize() const
Optimize this function for minimum size (-Oz).
Definition: Function.h:594
const MemCmpExpansionOptions * enableMemCmpExpansion(bool IsZeroCmp) const
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:322
uint32_t Size
Definition: Profile.cpp:46
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value *> Args=None, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:2009
BasicBlock * splitBasicBlock(iterator I, const Twine &BBName="")
Split the basic block into two basic blocks at the specified instruction.
Definition: BasicBlock.cpp:407
InstTy * Insert(InstTy *I, const Twine &Name="") const
Insert and return the specified instruction.
Definition: IRBuilder.h:793
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
virtual unsigned getMemcmpEqZeroLoadsPerBlock() const
For memcmp expansion when the memcmp result is only compared equal or not-equal to 0...
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:565
LLVM Value Representation.
Definition: Value.h:72
virtual bool runOnFunction(Function &F)=0
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass...
Primary interface to the complete machine description for the target machine.
Definition: TargetMachine.h:58
bool skipFunction(const Function &F) const
Optional passes call this function to check whether the pass should be skipped.
Definition: Pass.cpp:157
This pass exposes codegen information to IR-level passes.
hexagon widen stores
bool empty() const
empty - Check if the array is empty.
Definition: ArrayRef.h:143
This file describes how to lower LLVM code to machine code.
const BasicBlock * getParent() const
Definition: Instruction.h:66