lib/CodeGen/ExpandMemCmp.cpp

   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"
  15 #include "llvm/Analysis/ConstantFolding.h"
  16 #include "llvm/Analysis/TargetLibraryInfo.h"
  17 #include "llvm/Analysis/TargetTransformInfo.h"
  18 #include "llvm/Analysis/ValueTracking.h"
  19 #include "llvm/CodeGen/TargetLowering.h"
  20 #include "llvm/CodeGen/TargetPassConfig.h"
  21 #include "llvm/CodeGen/TargetSubtargetInfo.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
  34 static cl::opt<unsigned> MemCmpEqZeroNumLoadsPerBlock(
  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 static cl::opt<unsigned> MaxLoadsPerMemcmp(
  40     "max-loads-per-memcmp", cl::Hidden,
  41     cl::desc("Set maximum number of loads used in expanded memcmp"));
  42
  43 static cl::opt<unsigned> MaxLoadsPerMemcmpOptSize(
  44     "max-loads-per-memcmp-opt-size", cl::Hidden,
  45     cl::desc("Set maximum number of loads used in expanded memcmp for -Os/Oz"));
  46
  47 namespace {
  48
  49
  50 // This class provides helper functions to expand a memcmp library call into an
  51 // inline expansion.
  52 class MemCmpExpansion {
  53   struct ResultBlock {
  54     BasicBlock *BB = nullptr;
  55     PHINode *PhiSrc1 = nullptr;
  56     PHINode *PhiSrc2 = nullptr;
  57
  58     ResultBlock() = default;
  59   };
  60
  61   CallInst *const CI;
  62   ResultBlock ResBlock;
  63   const uint64_t Size;
  64   unsigned MaxLoadSize;
  65   uint64_t NumLoadsNonOneByte;
  66   const uint64_t NumLoadsPerBlockForZeroCmp;
  67   std::vector<BasicBlock *> LoadCmpBlocks;
  68   BasicBlock *EndBlock;
  69   PHINode *PhiRes;
  70   const bool IsUsedForZeroCmp;
  71   const DataLayout &DL;
  72   IRBuilder<> Builder;
  73   // Represents the decomposition in blocks of the expansion. For example,
  74   // comparing 33 bytes on X86+sse can be done with 2x16-byte loads and
  75   // 1x1-byte load, which would be represented as [{16, 0}, {16, 16}, {32, 1}.
  76   struct LoadEntry {
  77     LoadEntry(unsigned LoadSize, uint64_t Offset)
  78         : LoadSize(LoadSize), Offset(Offset) {
  79     }
  80
  81     // The size of the load for this block, in bytes.
  82     unsigned LoadSize;
  83     // The offset of this load from the base pointer, in bytes.
  84     uint64_t Offset;
  85   };
  86   using LoadEntryVector = SmallVector<LoadEntry, 8>;
  87   LoadEntryVector LoadSequence;
  88
  89   void createLoadCmpBlocks();
  90   void createResultBlock();
  91   void setupResultBlockPHINodes();
  92   void setupEndBlockPHINodes();
  93   Value *getCompareLoadPairs(unsigned BlockIndex, unsigned &LoadIndex);
  94   void emitLoadCompareBlock(unsigned BlockIndex);
  95   void emitLoadCompareBlockMultipleLoads(unsigned BlockIndex,
  96                                          unsigned &LoadIndex);
  97   void emitLoadCompareByteBlock(unsigned BlockIndex, unsigned OffsetBytes);
  98   void emitMemCmpResultBlock();
  99   Value *getMemCmpExpansionZeroCase();
 100   Value *getMemCmpEqZeroOneBlock();
 101   Value *getMemCmpOneBlock();
 102   Value *getPtrToElementAtOffset(Value *Source, Type *LoadSizeType,
 103                                  uint64_t OffsetBytes);
 104
 105   static LoadEntryVector
 106   computeGreedyLoadSequence(uint64_t Size, llvm::ArrayRef<unsigned> LoadSizes,
 107                             unsigned MaxNumLoads, unsigned &NumLoadsNonOneByte);
 108   static LoadEntryVector
 109   computeOverlappingLoadSequence(uint64_t Size, unsigned MaxLoadSize,
 110                                  unsigned MaxNumLoads,
 111                                  unsigned &NumLoadsNonOneByte);
 112
 113 public:
 114   MemCmpExpansion(CallInst *CI, uint64_t Size,
 115                   const TargetTransformInfo::MemCmpExpansionOptions &Options,
 116                   const bool IsUsedForZeroCmp, const DataLayout &TheDataLayout);
 117
 118   unsigned getNumBlocks();
 119   uint64_t getNumLoads() const { return LoadSequence.size(); }
 120
 121   Value *getMemCmpExpansion();
 122 };
 123
 124 MemCmpExpansion::LoadEntryVector MemCmpExpansion::computeGreedyLoadSequence(
 125     uint64_t Size, llvm::ArrayRef<unsigned> LoadSizes,
 126     const unsigned MaxNumLoads, unsigned &NumLoadsNonOneByte) {
 127   NumLoadsNonOneByte = 0;
 128   LoadEntryVector LoadSequence;
 129   uint64_t Offset = 0;
 130   while (Size && !LoadSizes.empty()) {
 131     const unsigned LoadSize = LoadSizes.front();
 132     const uint64_t NumLoadsForThisSize = Size / LoadSize;
 133     if (LoadSequence.size() + NumLoadsForThisSize > MaxNumLoads) {
 134       // Do not expand if the total number of loads is larger than what the
 135       // target allows. Note that it's important that we exit before completing
 136       // the expansion to avoid using a ton of memory to store the expansion for
 137       // large sizes.
 138       return {};
 139     }
 140     if (NumLoadsForThisSize > 0) {
 141       for (uint64_t I = 0; I < NumLoadsForThisSize; ++I) {
 142         LoadSequence.push_back({LoadSize, Offset});
 143         Offset += LoadSize;
 144       }
 145       if (LoadSize > 1)
 146         ++NumLoadsNonOneByte;
 147       Size = Size % LoadSize;
 148     }
 149     LoadSizes = LoadSizes.drop_front();
 150   }
 151   return LoadSequence;
 152 }
 153
 154 MemCmpExpansion::LoadEntryVector
 155 MemCmpExpansion::computeOverlappingLoadSequence(uint64_t Size,
 156                                                 const unsigned MaxLoadSize,
 157                                                 const unsigned MaxNumLoads,
 158                                                 unsigned &NumLoadsNonOneByte) {
 159   // These are already handled by the greedy approach.
 160   if (Size < 2 || MaxLoadSize < 2)
 161     return {};
 162
 163   // We try to do as many non-overlapping loads as possible starting from the
 164   // beginning.
 165   const uint64_t NumNonOverlappingLoads = Size / MaxLoadSize;
 166   assert(NumNonOverlappingLoads && "there must be at least one load");
 167   // There remain 0 to (MaxLoadSize - 1) bytes to load, this will be done with
 168   // an overlapping load.
 169   Size = Size - NumNonOverlappingLoads * MaxLoadSize;
 170   // Bail if we do not need an overloapping store, this is already handled by
 171   // the greedy approach.
 172   if (Size == 0)
 173     return {};
 174   // Bail if the number of loads (non-overlapping + potential overlapping one)
 175   // is larger than the max allowed.
 176   if ((NumNonOverlappingLoads + 1) > MaxNumLoads)
 177     return {};
 178
 179   // Add non-overlapping loads.
 180   LoadEntryVector LoadSequence;
 181   uint64_t Offset = 0;
 182   for (uint64_t I = 0; I < NumNonOverlappingLoads; ++I) {
 183     LoadSequence.push_back({MaxLoadSize, Offset});
 184     Offset += MaxLoadSize;
 185   }
 186
 187   // Add the last overlapping load.
 188   assert(Size > 0 && Size < MaxLoadSize && "broken invariant");
 189   LoadSequence.push_back({MaxLoadSize, Offset - (MaxLoadSize - Size)});
 190   NumLoadsNonOneByte = 1;
 191   return LoadSequence;
 192 }
 193
 194 // Initialize the basic block structure required for expansion of memcmp call
 195 // with given maximum load size and memcmp size parameter.
 196 // This structure includes:
 197 // 1. A list of load compare blocks - LoadCmpBlocks.
 198 // 2. An EndBlock, split from original instruction point, which is the block to
 199 // return from.
 200 // 3. ResultBlock, block to branch to for early exit when a
 201 // LoadCmpBlock finds a difference.
 202 MemCmpExpansion::MemCmpExpansion(
 203     CallInst *const CI, uint64_t Size,
 204     const TargetTransformInfo::MemCmpExpansionOptions &Options,
 205     const bool IsUsedForZeroCmp, const DataLayout &TheDataLayout)
 206     : CI(CI), Size(Size), MaxLoadSize(0), NumLoadsNonOneByte(0),
 207       NumLoadsPerBlockForZeroCmp(Options.NumLoadsPerBlock),
 208       IsUsedForZeroCmp(IsUsedForZeroCmp), DL(TheDataLayout), Builder(CI) {
 209   assert(Size > 0 && "zero blocks");
 210   // Scale the max size down if the target can load more bytes than we need.
 211   llvm::ArrayRef<unsigned> LoadSizes(Options.LoadSizes);
 212   while (!LoadSizes.empty() && LoadSizes.front() > Size) {
 213     LoadSizes = LoadSizes.drop_front();
 214   }
 215   assert(!LoadSizes.empty() && "cannot load Size bytes");
 216   MaxLoadSize = LoadSizes.front();
 217   // Compute the decomposition.
 218   unsigned GreedyNumLoadsNonOneByte = 0;
 219   LoadSequence = computeGreedyLoadSequence(Size, LoadSizes, Options.MaxNumLoads,
 220                                            GreedyNumLoadsNonOneByte);
 221   NumLoadsNonOneByte = GreedyNumLoadsNonOneByte;
 222   assert(LoadSequence.size() <= Options.MaxNumLoads && "broken invariant");
 223   // If we allow overlapping loads and the load sequence is not already optimal,
 224   // use overlapping loads.
 225   if (Options.AllowOverlappingLoads &&
 226       (LoadSequence.empty() || LoadSequence.size() > 2)) {
 227     unsigned OverlappingNumLoadsNonOneByte = 0;
 228     auto OverlappingLoads = computeOverlappingLoadSequence(
 229         Size, MaxLoadSize, Options.MaxNumLoads, OverlappingNumLoadsNonOneByte);
 230     if (!OverlappingLoads.empty() &&
 231         (LoadSequence.empty() ||
 232          OverlappingLoads.size() < LoadSequence.size())) {
 233       LoadSequence = OverlappingLoads;
 234       NumLoadsNonOneByte = OverlappingNumLoadsNonOneByte;
 235     }
 236   }
 237   assert(LoadSequence.size() <= Options.MaxNumLoads && "broken invariant");
 238 }
 239
 240 unsigned MemCmpExpansion::getNumBlocks() {
 241   if (IsUsedForZeroCmp)
 242     return getNumLoads() / NumLoadsPerBlockForZeroCmp +
 243            (getNumLoads() % NumLoadsPerBlockForZeroCmp != 0 ? 1 : 0);
 244   return getNumLoads();
 245 }
 246
 247 void MemCmpExpansion::createLoadCmpBlocks() {
 248   for (unsigned i = 0; i < getNumBlocks(); i++) {
 249     BasicBlock *BB = BasicBlock::Create(CI->getContext(), "loadbb",
 250                                         EndBlock->getParent(), EndBlock);
 251     LoadCmpBlocks.push_back(BB);
 252   }
 253 }
 254
 255 void MemCmpExpansion::createResultBlock() {
 256   ResBlock.BB = BasicBlock::Create(CI->getContext(), "res_block",
 257                                    EndBlock->getParent(), EndBlock);
 258 }
 259
 260 /// Return a pointer to an element of type `LoadSizeType` at offset
 261 /// `OffsetBytes`.
 262 Value *MemCmpExpansion::getPtrToElementAtOffset(Value *Source,
 263                                                 Type *LoadSizeType,
 264                                                 uint64_t OffsetBytes) {
 265   if (OffsetBytes > 0) {
 266     auto *ByteType = Type::getInt8Ty(CI->getContext());
 267     Source = Builder.CreateGEP(
 268         ByteType, Builder.CreateBitCast(Source, ByteType->getPointerTo()),
 269         ConstantInt::get(ByteType, OffsetBytes));
 270   }
 271   return Builder.CreateBitCast(Source, LoadSizeType->getPointerTo());
 272 }
 273
 274 // This function creates the IR instructions for loading and comparing 1 byte.
 275 // It loads 1 byte from each source of the memcmp parameters with the given
 276 // GEPIndex. It then subtracts the two loaded values and adds this result to the
 277 // final phi node for selecting the memcmp result.
 278 void MemCmpExpansion::emitLoadCompareByteBlock(unsigned BlockIndex,
 279                                                unsigned OffsetBytes) {
 280   Builder.SetInsertPoint(LoadCmpBlocks[BlockIndex]);
 281   Type *LoadSizeType = Type::getInt8Ty(CI->getContext());
 282   Value *Source1 =
 283       getPtrToElementAtOffset(CI->getArgOperand(0), LoadSizeType, OffsetBytes);
 284   Value *Source2 =
 285       getPtrToElementAtOffset(CI->getArgOperand(1), LoadSizeType, OffsetBytes);
 286
 287   Value *LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
 288   Value *LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
 289
 290   LoadSrc1 = Builder.CreateZExt(LoadSrc1, Type::getInt32Ty(CI->getContext()));
 291   LoadSrc2 = Builder.CreateZExt(LoadSrc2, Type::getInt32Ty(CI->getContext()));
 292   Value *Diff = Builder.CreateSub(LoadSrc1, LoadSrc2);
 293
 294   PhiRes->addIncoming(Diff, LoadCmpBlocks[BlockIndex]);
 295
 296   if (BlockIndex < (LoadCmpBlocks.size() - 1)) {
 297     // Early exit branch if difference found to EndBlock. Otherwise, continue to
 298     // next LoadCmpBlock,
 299     Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_NE, Diff,
 300                                     ConstantInt::get(Diff->getType(), 0));
 301     BranchInst *CmpBr =
 302         BranchInst::Create(EndBlock, LoadCmpBlocks[BlockIndex + 1], Cmp);
 303     Builder.Insert(CmpBr);
 304   } else {
 305     // The last block has an unconditional branch to EndBlock.
 306     BranchInst *CmpBr = BranchInst::Create(EndBlock);
 307     Builder.Insert(CmpBr);
 308   }
 309 }
 310
 311 /// Generate an equality comparison for one or more pairs of loaded values.
 312 /// This is used in the case where the memcmp() call is compared equal or not
 313 /// equal to zero.
 314 Value *MemCmpExpansion::getCompareLoadPairs(unsigned BlockIndex,
 315                                             unsigned &LoadIndex) {
 316   assert(LoadIndex < getNumLoads() &&
 317          "getCompareLoadPairs() called with no remaining loads");
 318   std::vector<Value *> XorList, OrList;
 319   Value *Diff = nullptr;
 320
 321   const unsigned NumLoads =
 322       std::min(getNumLoads() - LoadIndex, NumLoadsPerBlockForZeroCmp);
 323
 324   // For a single-block expansion, start inserting before the memcmp call.
 325   if (LoadCmpBlocks.empty())
 326     Builder.SetInsertPoint(CI);
 327   else
 328     Builder.SetInsertPoint(LoadCmpBlocks[BlockIndex]);
 329
 330   Value *Cmp = nullptr;
 331   // If we have multiple loads per block, we need to generate a composite
 332   // comparison using xor+or. The type for the combinations is the largest load
 333   // type.
 334   IntegerType *const MaxLoadType =
 335       NumLoads == 1 ? nullptr
 336                     : IntegerType::get(CI->getContext(), MaxLoadSize * 8);
 337   for (unsigned i = 0; i < NumLoads; ++i, ++LoadIndex) {
 338     const LoadEntry &CurLoadEntry = LoadSequence[LoadIndex];
 339
 340     IntegerType *LoadSizeType =
 341         IntegerType::get(CI->getContext(), CurLoadEntry.LoadSize * 8);
 342
 343     Value *Source1 = getPtrToElementAtOffset(CI->getArgOperand(0), LoadSizeType,
 344                                              CurLoadEntry.Offset);
 345     Value *Source2 = getPtrToElementAtOffset(CI->getArgOperand(1), LoadSizeType,
 346                                              CurLoadEntry.Offset);
 347
 348     // Get a constant or load a value for each source address.
 349     Value *LoadSrc1 = nullptr;
 350     if (auto *Source1C = dyn_cast<Constant>(Source1))
 351       LoadSrc1 = ConstantFoldLoadFromConstPtr(Source1C, LoadSizeType, DL);
 352     if (!LoadSrc1)
 353       LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
 354
 355     Value *LoadSrc2 = nullptr;
 356     if (auto *Source2C = dyn_cast<Constant>(Source2))
 357       LoadSrc2 = ConstantFoldLoadFromConstPtr(Source2C, LoadSizeType, DL);
 358     if (!LoadSrc2)
 359       LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
 360
 361     if (NumLoads != 1) {
 362       if (LoadSizeType != MaxLoadType) {
 363         LoadSrc1 = Builder.CreateZExt(LoadSrc1, MaxLoadType);
 364         LoadSrc2 = Builder.CreateZExt(LoadSrc2, MaxLoadType);
 365       }
 366       // If we have multiple loads per block, we need to generate a composite
 367       // comparison using xor+or.
 368       Diff = Builder.CreateXor(LoadSrc1, LoadSrc2);
 369       Diff = Builder.CreateZExt(Diff, MaxLoadType);
 370       XorList.push_back(Diff);
 371     } else {
 372       // If there's only one load per block, we just compare the loaded values.
 373       Cmp = Builder.CreateICmpNE(LoadSrc1, LoadSrc2);
 374     }
 375   }
 376
 377   auto pairWiseOr = [&](std::vector<Value *> &InList) -> std::vector<Value *> {
 378     std::vector<Value *> OutList;
 379     for (unsigned i = 0; i < InList.size() - 1; i = i + 2) {
 380       Value *Or = Builder.CreateOr(InList[i], InList[i + 1]);
 381       OutList.push_back(Or);
 382     }
 383     if (InList.size() % 2 != 0)
 384       OutList.push_back(InList.back());
 385     return OutList;
 386   };
 387
 388   if (!Cmp) {
 389     // Pairwise OR the XOR results.
 390     OrList = pairWiseOr(XorList);
 391
 392     // Pairwise OR the OR results until one result left.
 393     while (OrList.size() != 1) {
 394       OrList = pairWiseOr(OrList);
 395     }
 396
 397     assert(Diff && "Failed to find comparison diff");
 398     Cmp = Builder.CreateICmpNE(OrList[0], ConstantInt::get(Diff->getType(), 0));
 399   }
 400
 401   return Cmp;
 402 }
 403
 404 void MemCmpExpansion::emitLoadCompareBlockMultipleLoads(unsigned BlockIndex,
 405                                                         unsigned &LoadIndex) {
 406   Value *Cmp = getCompareLoadPairs(BlockIndex, LoadIndex);
 407
 408   BasicBlock *NextBB = (BlockIndex == (LoadCmpBlocks.size() - 1))
 409                            ? EndBlock
 410                            : LoadCmpBlocks[BlockIndex + 1];
 411   // Early exit branch if difference found to ResultBlock. Otherwise,
 412   // continue to next LoadCmpBlock or EndBlock.
 413   BranchInst *CmpBr = BranchInst::Create(ResBlock.BB, NextBB, Cmp);
 414   Builder.Insert(CmpBr);
 415
 416   // Add a phi edge for the last LoadCmpBlock to Endblock with a value of 0
 417   // since early exit to ResultBlock was not taken (no difference was found in
 418   // any of the bytes).
 419   if (BlockIndex == LoadCmpBlocks.size() - 1) {
 420     Value *Zero = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 0);
 421     PhiRes->addIncoming(Zero, LoadCmpBlocks[BlockIndex]);
 422   }
 423 }
 424
 425 // This function creates the IR intructions for loading and comparing using the
 426 // given LoadSize. It loads the number of bytes specified by LoadSize from each
 427 // source of the memcmp parameters. It then does a subtract to see if there was
 428 // a difference in the loaded values. If a difference is found, it branches
 429 // with an early exit to the ResultBlock for calculating which source was
 430 // larger. Otherwise, it falls through to the either the next LoadCmpBlock or
 431 // the EndBlock if this is the last LoadCmpBlock. Loading 1 byte is handled with
 432 // a special case through emitLoadCompareByteBlock. The special handling can
 433 // simply subtract the loaded values and add it to the result phi node.
 434 void MemCmpExpansion::emitLoadCompareBlock(unsigned BlockIndex) {
 435   // There is one load per block in this case, BlockIndex == LoadIndex.
 436   const LoadEntry &CurLoadEntry = LoadSequence[BlockIndex];
 437
 438   if (CurLoadEntry.LoadSize == 1) {
 439     MemCmpExpansion::emitLoadCompareByteBlock(BlockIndex, CurLoadEntry.Offset);
 440     return;
 441   }
 442
 443   Type *LoadSizeType =
 444       IntegerType::get(CI->getContext(), CurLoadEntry.LoadSize * 8);
 445   Type *MaxLoadType = IntegerType::get(CI->getContext(), MaxLoadSize * 8);
 446   assert(CurLoadEntry.LoadSize <= MaxLoadSize && "Unexpected load type");
 447
 448   Builder.SetInsertPoint(LoadCmpBlocks[BlockIndex]);
 449
 450   Value *Source1 = getPtrToElementAtOffset(CI->getArgOperand(0), LoadSizeType,
 451                                            CurLoadEntry.Offset);
 452   Value *Source2 = getPtrToElementAtOffset(CI->getArgOperand(1), LoadSizeType,
 453                                            CurLoadEntry.Offset);
 454
 455   // Load LoadSizeType from the base address.
 456   Value *LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
 457   Value *LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
 458
 459   if (DL.isLittleEndian()) {
 460     Function *Bswap = Intrinsic::getDeclaration(CI->getModule(),
 461                                                 Intrinsic::bswap, LoadSizeType);
 462     LoadSrc1 = Builder.CreateCall(Bswap, LoadSrc1);
 463     LoadSrc2 = Builder.CreateCall(Bswap, LoadSrc2);
 464   }
 465
 466   if (LoadSizeType != MaxLoadType) {
 467     LoadSrc1 = Builder.CreateZExt(LoadSrc1, MaxLoadType);
 468     LoadSrc2 = Builder.CreateZExt(LoadSrc2, MaxLoadType);
 469   }
 470
 471   // Add the loaded values to the phi nodes for calculating memcmp result only
 472   // if result is not used in a zero equality.
 473   if (!IsUsedForZeroCmp) {
 474     ResBlock.PhiSrc1->addIncoming(LoadSrc1, LoadCmpBlocks[BlockIndex]);
 475     ResBlock.PhiSrc2->addIncoming(LoadSrc2, LoadCmpBlocks[BlockIndex]);
 476   }
 477
 478   Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, LoadSrc1, LoadSrc2);
 479   BasicBlock *NextBB = (BlockIndex == (LoadCmpBlocks.size() - 1))
 480                            ? EndBlock
 481                            : LoadCmpBlocks[BlockIndex + 1];
 482   // Early exit branch if difference found to ResultBlock. Otherwise, continue
 483   // to next LoadCmpBlock or EndBlock.
 484   BranchInst *CmpBr = BranchInst::Create(NextBB, ResBlock.BB, Cmp);
 485   Builder.Insert(CmpBr);
 486
 487   // Add a phi edge for the last LoadCmpBlock to Endblock with a value of 0
 488   // since early exit to ResultBlock was not taken (no difference was found in
 489   // any of the bytes).
 490   if (BlockIndex == LoadCmpBlocks.size() - 1) {
 491     Value *Zero = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 0);
 492     PhiRes->addIncoming(Zero, LoadCmpBlocks[BlockIndex]);
 493   }
 494 }
 495
 496 // This function populates the ResultBlock with a sequence to calculate the
 497 // memcmp result. It compares the two loaded source values and returns -1 if
 498 // src1 < src2 and 1 if src1 > src2.
 499 void MemCmpExpansion::emitMemCmpResultBlock() {
 500   // Special case: if memcmp result is used in a zero equality, result does not
 501   // need to be calculated and can simply return 1.
 502   if (IsUsedForZeroCmp) {
 503     BasicBlock::iterator InsertPt = ResBlock.BB->getFirstInsertionPt();
 504     Builder.SetInsertPoint(ResBlock.BB, InsertPt);
 505     Value *Res = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 1);
 506     PhiRes->addIncoming(Res, ResBlock.BB);
 507     BranchInst *NewBr = BranchInst::Create(EndBlock);
 508     Builder.Insert(NewBr);
 509     return;
 510   }
 511   BasicBlock::iterator InsertPt = ResBlock.BB->getFirstInsertionPt();
 512   Builder.SetInsertPoint(ResBlock.BB, InsertPt);
 513
 514   Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_ULT, ResBlock.PhiSrc1,
 515                                   ResBlock.PhiSrc2);
 516
 517   Value *Res =
 518       Builder.CreateSelect(Cmp, ConstantInt::get(Builder.getInt32Ty(), -1),
 519                            ConstantInt::get(Builder.getInt32Ty(), 1));
 520
 521   BranchInst *NewBr = BranchInst::Create(EndBlock);
 522   Builder.Insert(NewBr);
 523   PhiRes->addIncoming(Res, ResBlock.BB);
 524 }
 525
 526 void MemCmpExpansion::setupResultBlockPHINodes() {
 527   Type *MaxLoadType = IntegerType::get(CI->getContext(), MaxLoadSize * 8);
 528   Builder.SetInsertPoint(ResBlock.BB);
 529   // Note: this assumes one load per block.
 530   ResBlock.PhiSrc1 =
 531       Builder.CreatePHI(MaxLoadType, NumLoadsNonOneByte, "phi.src1");
 532   ResBlock.PhiSrc2 =
 533       Builder.CreatePHI(MaxLoadType, NumLoadsNonOneByte, "phi.src2");
 534 }
 535
 536 void MemCmpExpansion::setupEndBlockPHINodes() {
 537   Builder.SetInsertPoint(&EndBlock->front());
 538   PhiRes = Builder.CreatePHI(Type::getInt32Ty(CI->getContext()), 2, "phi.res");
 539 }
 540
 541 Value *MemCmpExpansion::getMemCmpExpansionZeroCase() {
 542   unsigned LoadIndex = 0;
 543   // This loop populates each of the LoadCmpBlocks with the IR sequence to
 544   // handle multiple loads per block.
 545   for (unsigned I = 0; I < getNumBlocks(); ++I) {
 546     emitLoadCompareBlockMultipleLoads(I, LoadIndex);
 547   }
 548
 549   emitMemCmpResultBlock();
 550   return PhiRes;
 551 }
 552
 553 /// A memcmp expansion that compares equality with 0 and only has one block of
 554 /// load and compare can bypass the compare, branch, and phi IR that is required
 555 /// in the general case.
 556 Value *MemCmpExpansion::getMemCmpEqZeroOneBlock() {
 557   unsigned LoadIndex = 0;
 558   Value *Cmp = getCompareLoadPairs(0, LoadIndex);
 559   assert(LoadIndex == getNumLoads() && "some entries were not consumed");
 560   return Builder.CreateZExt(Cmp, Type::getInt32Ty(CI->getContext()));
 561 }
 562
 563 /// A memcmp expansion that only has one block of load and compare can bypass
 564 /// the compare, branch, and phi IR that is required in the general case.
 565 Value *MemCmpExpansion::getMemCmpOneBlock() {
 566   Type *LoadSizeType = IntegerType::get(CI->getContext(), Size * 8);
 567   Value *Source1 = CI->getArgOperand(0);
 568   Value *Source2 = CI->getArgOperand(1);
 569
 570   // Cast source to LoadSizeType*.
 571   if (Source1->getType() != LoadSizeType)
 572     Source1 = Builder.CreateBitCast(Source1, LoadSizeType->getPointerTo());
 573   if (Source2->getType() != LoadSizeType)
 574     Source2 = Builder.CreateBitCast(Source2, LoadSizeType->getPointerTo());
 575
 576   // Load LoadSizeType from the base address.
 577   Value *LoadSrc1 = Builder.CreateLoad(LoadSizeType, Source1);
 578   Value *LoadSrc2 = Builder.CreateLoad(LoadSizeType, Source2);
 579
 580   if (DL.isLittleEndian() && Size != 1) {
 581     Function *Bswap = Intrinsic::getDeclaration(CI->getModule(),
 582                                                 Intrinsic::bswap, LoadSizeType);
 583     LoadSrc1 = Builder.CreateCall(Bswap, LoadSrc1);
 584     LoadSrc2 = Builder.CreateCall(Bswap, LoadSrc2);
 585   }
 586
 587   if (Size < 4) {
 588     // The i8 and i16 cases don't need compares. We zext the loaded values and
 589     // subtract them to get the suitable negative, zero, or positive i32 result.
 590     LoadSrc1 = Builder.CreateZExt(LoadSrc1, Builder.getInt32Ty());
 591     LoadSrc2 = Builder.CreateZExt(LoadSrc2, Builder.getInt32Ty());
 592     return Builder.CreateSub(LoadSrc1, LoadSrc2);
 593   }
 594
 595   // The result of memcmp is negative, zero, or positive, so produce that by
 596   // subtracting 2 extended compare bits: sub (ugt, ult).
 597   // If a target prefers to use selects to get -1/0/1, they should be able
 598   // to transform this later. The inverse transform (going from selects to math)
 599   // may not be possible in the DAG because the selects got converted into
 600   // branches before we got there.
 601   Value *CmpUGT = Builder.CreateICmpUGT(LoadSrc1, LoadSrc2);
 602   Value *CmpULT = Builder.CreateICmpULT(LoadSrc1, LoadSrc2);
 603   Value *ZextUGT = Builder.CreateZExt(CmpUGT, Builder.getInt32Ty());
 604   Value *ZextULT = Builder.CreateZExt(CmpULT, Builder.getInt32Ty());
 605   return Builder.CreateSub(ZextUGT, ZextULT);
 606 }
 607
 608 // This function expands the memcmp call into an inline expansion and returns
 609 // the memcmp result.
 610 Value *MemCmpExpansion::getMemCmpExpansion() {
 611   // Create the basic block framework for a multi-block expansion.
 612   if (getNumBlocks() != 1) {
 613     BasicBlock *StartBlock = CI->getParent();
 614     EndBlock = StartBlock->splitBasicBlock(CI, "endblock");
 615     setupEndBlockPHINodes();
 616     createResultBlock();
 617
 618     // If return value of memcmp is not used in a zero equality, we need to
 619     // calculate which source was larger. The calculation requires the
 620     // two loaded source values of each load compare block.
 621     // These will be saved in the phi nodes created by setupResultBlockPHINodes.
 622     if (!IsUsedForZeroCmp) setupResultBlockPHINodes();
 623
 624     // Create the number of required load compare basic blocks.
 625     createLoadCmpBlocks();
 626
 627     // Update the terminator added by splitBasicBlock to branch to the first
 628     // LoadCmpBlock.
 629     StartBlock->getTerminator()->setSuccessor(0, LoadCmpBlocks[0]);
 630   }
 631
 632   Builder.SetCurrentDebugLocation(CI->getDebugLoc());
 633
 634   if (IsUsedForZeroCmp)
 635     return getNumBlocks() == 1 ? getMemCmpEqZeroOneBlock()
 636                                : getMemCmpExpansionZeroCase();
 637
 638   if (getNumBlocks() == 1)
 639     return getMemCmpOneBlock();
 640
 641   for (unsigned I = 0; I < getNumBlocks(); ++I) {
 642     emitLoadCompareBlock(I);
 643   }
 644
 645   emitMemCmpResultBlock();
 646   return PhiRes;
 647 }
 648
 649 // This function checks to see if an expansion of memcmp can be generated.
 650 // It checks for constant compare size that is less than the max inline size.
 651 // If an expansion cannot occur, returns false to leave as a library call.
 652 // Otherwise, the library call is replaced with a new IR instruction sequence.
 653 /// We want to transform:
 654 /// %call = call signext i32 @memcmp(i8* %0, i8* %1, i64 15)
 655 /// To:
 656 /// loadbb:
 657 ///  %0 = bitcast i32* %buffer2 to i8*
 658 ///  %1 = bitcast i32* %buffer1 to i8*
 659 ///  %2 = bitcast i8* %1 to i64*
 660 ///  %3 = bitcast i8* %0 to i64*
 661 ///  %4 = load i64, i64* %2
 662 ///  %5 = load i64, i64* %3
 663 ///  %6 = call i64 @llvm.bswap.i64(i64 %4)
 664 ///  %7 = call i64 @llvm.bswap.i64(i64 %5)
 665 ///  %8 = sub i64 %6, %7
 666 ///  %9 = icmp ne i64 %8, 0
 667 ///  br i1 %9, label %res_block, label %loadbb1
 668 /// res_block:                                        ; preds = %loadbb2,
 669 /// %loadbb1, %loadbb
 670 ///  %phi.src1 = phi i64 [ %6, %loadbb ], [ %22, %loadbb1 ], [ %36, %loadbb2 ]
 671 ///  %phi.src2 = phi i64 [ %7, %loadbb ], [ %23, %loadbb1 ], [ %37, %loadbb2 ]
 672 ///  %10 = icmp ult i64 %phi.src1, %phi.src2
 673 ///  %11 = select i1 %10, i32 -1, i32 1
 674 ///  br label %endblock
 675 /// loadbb1:                                          ; preds = %loadbb
 676 ///  %12 = bitcast i32* %buffer2 to i8*
 677 ///  %13 = bitcast i32* %buffer1 to i8*
 678 ///  %14 = bitcast i8* %13 to i32*
 679 ///  %15 = bitcast i8* %12 to i32*
 680 ///  %16 = getelementptr i32, i32* %14, i32 2
 681 ///  %17 = getelementptr i32, i32* %15, i32 2
 682 ///  %18 = load i32, i32* %16
 683 ///  %19 = load i32, i32* %17
 684 ///  %20 = call i32 @llvm.bswap.i32(i32 %18)
 685 ///  %21 = call i32 @llvm.bswap.i32(i32 %19)
 686 ///  %22 = zext i32 %20 to i64
 687 ///  %23 = zext i32 %21 to i64
 688 ///  %24 = sub i64 %22, %23
 689 ///  %25 = icmp ne i64 %24, 0
 690 ///  br i1 %25, label %res_block, label %loadbb2
 691 /// loadbb2:                                          ; preds = %loadbb1
 692 ///  %26 = bitcast i32* %buffer2 to i8*
 693 ///  %27 = bitcast i32* %buffer1 to i8*
 694 ///  %28 = bitcast i8* %27 to i16*
 695 ///  %29 = bitcast i8* %26 to i16*
 696 ///  %30 = getelementptr i16, i16* %28, i16 6
 697 ///  %31 = getelementptr i16, i16* %29, i16 6
 698 ///  %32 = load i16, i16* %30
 699 ///  %33 = load i16, i16* %31
 700 ///  %34 = call i16 @llvm.bswap.i16(i16 %32)
 701 ///  %35 = call i16 @llvm.bswap.i16(i16 %33)
 702 ///  %36 = zext i16 %34 to i64
 703 ///  %37 = zext i16 %35 to i64
 704 ///  %38 = sub i64 %36, %37
 705 ///  %39 = icmp ne i64 %38, 0
 706 ///  br i1 %39, label %res_block, label %loadbb3
 707 /// loadbb3:                                          ; preds = %loadbb2
 708 ///  %40 = bitcast i32* %buffer2 to i8*
 709 ///  %41 = bitcast i32* %buffer1 to i8*
 710 ///  %42 = getelementptr i8, i8* %41, i8 14
 711 ///  %43 = getelementptr i8, i8* %40, i8 14
 712 ///  %44 = load i8, i8* %42
 713 ///  %45 = load i8, i8* %43
 714 ///  %46 = zext i8 %44 to i32
 715 ///  %47 = zext i8 %45 to i32
 716 ///  %48 = sub i32 %46, %47
 717 ///  br label %endblock
 718 /// endblock:                                         ; preds = %res_block,
 719 /// %loadbb3
 720 ///  %phi.res = phi i32 [ %48, %loadbb3 ], [ %11, %res_block ]
 721 ///  ret i32 %phi.res
 722 static bool expandMemCmp(CallInst *CI, const TargetTransformInfo *TTI,
 723                          const TargetLowering *TLI, const DataLayout *DL) {
 724   NumMemCmpCalls++;
 725
 726   // Early exit from expansion if -Oz.
 727   if (CI->getFunction()->hasMinSize())
 728     return false;
 729
 730   // Early exit from expansion if size is not a constant.
 731   ConstantInt *SizeCast = dyn_cast<ConstantInt>(CI->getArgOperand(2));
 732   if (!SizeCast) {
 733     NumMemCmpNotConstant++;
 734     return false;
 735   }
 736   const uint64_t SizeVal = SizeCast->getZExtValue();
 737
 738   if (SizeVal == 0) {
 739     return false;
 740   }
 741   // TTI call to check if target would like to expand memcmp. Also, get the
 742   // available load sizes.
 743   const bool IsUsedForZeroCmp = isOnlyUsedInZeroEqualityComparison(CI);
 744   auto Options = TTI->enableMemCmpExpansion(CI->getFunction()->hasOptSize(),
 745                                             IsUsedForZeroCmp);
 746   if (!Options) return false;
 747
 748   if (MemCmpEqZeroNumLoadsPerBlock.getNumOccurrences())
 749     Options.NumLoadsPerBlock = MemCmpEqZeroNumLoadsPerBlock;
 750
 751   if (CI->getFunction()->hasOptSize() &&
 752       MaxLoadsPerMemcmpOptSize.getNumOccurrences())
 753     Options.MaxNumLoads = MaxLoadsPerMemcmpOptSize;
 754
 755   if (!CI->getFunction()->hasOptSize() && MaxLoadsPerMemcmp.getNumOccurrences())
 756     Options.MaxNumLoads = MaxLoadsPerMemcmp;
 757
 758   MemCmpExpansion Expansion(CI, SizeVal, Options, IsUsedForZeroCmp, *DL);
 759
 760   // Don't expand if this will require more loads than desired by the target.
 761   if (Expansion.getNumLoads() == 0) {
 762     NumMemCmpGreaterThanMax++;
 763     return false;
 764   }
 765
 766   NumMemCmpInlined++;
 767
 768   Value *Res = Expansion.getMemCmpExpansion();
 769
 770   // Replace call with result of expansion and erase call.
 771   CI->replaceAllUsesWith(Res);
 772   CI->eraseFromParent();
 773
 774   return true;
 775 }
 776
 777
 778
 779 class ExpandMemCmpPass : public FunctionPass {
 780 public:
 781   static char ID;
 782
 783   ExpandMemCmpPass() : FunctionPass(ID) {
 784     initializeExpandMemCmpPassPass(*PassRegistry::getPassRegistry());
 785   }
 786
 787   bool runOnFunction(Function &F) override {
 788     if (skipFunction(F)) return false;
 789
 790     auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
 791     if (!TPC) {
 792       return false;
 793     }
 794     const TargetLowering* TL =
 795         TPC->getTM<TargetMachine>().getSubtargetImpl(F)->getTargetLowering();
 796
 797     const TargetLibraryInfo *TLI =
 798         &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
 799     const TargetTransformInfo *TTI =
 800         &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
 801     auto PA = runImpl(F, TLI, TTI, TL);
 802     return !PA.areAllPreserved();
 803   }
 804
 805 private:
 806   void getAnalysisUsage(AnalysisUsage &AU) const override {
 807     AU.addRequired<TargetLibraryInfoWrapperPass>();
 808     AU.addRequired<TargetTransformInfoWrapperPass>();
 809     FunctionPass::getAnalysisUsage(AU);
 810   }
 811
 812   PreservedAnalyses runImpl(Function &F, const TargetLibraryInfo *TLI,
 813                             const TargetTransformInfo *TTI,
 814                             const TargetLowering* TL);
 815   // Returns true if a change was made.
 816   bool runOnBlock(BasicBlock &BB, const TargetLibraryInfo *TLI,
 817                   const TargetTransformInfo *TTI, const TargetLowering* TL,
 818                   const DataLayout& DL);
 819 };
 820
 821 bool ExpandMemCmpPass::runOnBlock(
 822     BasicBlock &BB, const TargetLibraryInfo *TLI,
 823     const TargetTransformInfo *TTI, const TargetLowering* TL,
 824     const DataLayout& DL) {
 825   for (Instruction& I : BB) {
 826     CallInst *CI = dyn_cast<CallInst>(&I);
 827     if (!CI) {
 828       continue;
 829     }
 830     LibFunc Func;
 831     if (TLI->getLibFunc(ImmutableCallSite(CI), Func) &&
 832         (Func == LibFunc_memcmp || Func == LibFunc_bcmp) &&
 833         expandMemCmp(CI, TTI, TL, &DL)) {
 834       return true;
 835     }
 836   }
 837   return false;
 838 }
 839
 840
 841 PreservedAnalyses ExpandMemCmpPass::runImpl(
 842     Function &F, const TargetLibraryInfo *TLI, const TargetTransformInfo *TTI,
 843     const TargetLowering* TL) {
 844   const DataLayout& DL = F.getParent()->getDataLayout();
 845   bool MadeChanges = false;
 846   for (auto BBIt = F.begin(); BBIt != F.end();) {
 847     if (runOnBlock(*BBIt, TLI, TTI, TL, DL)) {
 848       MadeChanges = true;
 849       // If changes were made, restart the function from the beginning, since
 850       // the structure of the function was changed.
 851       BBIt = F.begin();
 852     } else {
 853       ++BBIt;
 854     }
 855   }
 856   return MadeChanges ? PreservedAnalyses::none() : PreservedAnalyses::all();
 857 }
 858
 859 } // namespace
 860
 861 char ExpandMemCmpPass::ID = 0;
 862 INITIALIZE_PASS_BEGIN(ExpandMemCmpPass, "expandmemcmp",
 863                       "Expand memcmp() to load/stores", false, false)
 864 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
 865 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
 866 INITIALIZE_PASS_END(ExpandMemCmpPass, "expandmemcmp",
 867                     "Expand memcmp() to load/stores", false, false)
 868
 869 FunctionPass *llvm::createExpandMemCmpPass() {
 870   return new ExpandMemCmpPass();
 871 }