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/DomTreeUpdater.h"
17 #include "llvm/Analysis/LazyBlockFrequencyInfo.h"
18 #include "llvm/Analysis/ProfileSummaryInfo.h"
19 #include "llvm/Analysis/TargetLibraryInfo.h"
20 #include "llvm/Analysis/TargetTransformInfo.h"
21 #include "llvm/Analysis/ValueTracking.h"
22 #include "llvm/CodeGen/TargetLowering.h"
23 #include "llvm/CodeGen/TargetPassConfig.h"
24 #include "llvm/CodeGen/TargetSubtargetInfo.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/InitializePasses.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 #include "llvm/Transforms/Utils/SizeOpts.h"
32 
33 using namespace llvm;
34 
35 #define DEBUG_TYPE "expandmemcmp"
36 
37 STATISTIC(NumMemCmpCalls, "Number of memcmp calls");
38 STATISTIC(NumMemCmpNotConstant, "Number of memcmp calls without constant size");
39 STATISTIC(NumMemCmpGreaterThanMax,
40           "Number of memcmp calls with size greater than max size");
41 STATISTIC(NumMemCmpInlined, "Number of inlined memcmp calls");
42 
43 static cl::opt<unsigned> MemCmpEqZeroNumLoadsPerBlock(
44     "memcmp-num-loads-per-block", cl::Hidden, cl::init(1),
45     cl::desc("The number of loads per basic block for inline expansion of "
46              "memcmp that is only being compared against zero."));
47 
48 static cl::opt<unsigned> MaxLoadsPerMemcmp(
49     "max-loads-per-memcmp", cl::Hidden,
50     cl::desc("Set maximum number of loads used in expanded memcmp"));
51 
52 static cl::opt<unsigned> MaxLoadsPerMemcmpOptSize(
53     "max-loads-per-memcmp-opt-size", cl::Hidden,
54     cl::desc("Set maximum number of loads used in expanded memcmp for -Os/Oz"));
55 
56 namespace {
57 
58 
59 // This class provides helper functions to expand a memcmp library call into an
60 // inline expansion.
61 class MemCmpExpansion {
62   struct ResultBlock {
63     BasicBlock *BB = nullptr;
64     PHINode *PhiSrc1 = nullptr;
65     PHINode *PhiSrc2 = nullptr;
66 
67     ResultBlock() = default;
68   };
69 
70   CallInst *const CI;
71   ResultBlock ResBlock;
72   const uint64_t Size;
73   unsigned MaxLoadSize;
74   uint64_t NumLoadsNonOneByte;
75   const uint64_t NumLoadsPerBlockForZeroCmp;
76   std::vector<BasicBlock *> LoadCmpBlocks;
77   BasicBlock *EndBlock;
78   PHINode *PhiRes;
79   const bool IsUsedForZeroCmp;
80   const DataLayout &DL;
81   DomTreeUpdater *DTU;
82   IRBuilder<> Builder;
83   // Represents the decomposition in blocks of the expansion. For example,
84   // comparing 33 bytes on X86+sse can be done with 2x16-byte loads and
85   // 1x1-byte load, which would be represented as [{16, 0}, {16, 16}, {1, 32}.
86   struct LoadEntry {
LoadEntry__anon33d6824d0111::MemCmpExpansion::LoadEntry87     LoadEntry(unsigned LoadSize, uint64_t Offset)
88         : LoadSize(LoadSize), Offset(Offset) {
89     }
90 
91     // The size of the load for this block, in bytes.
92     unsigned LoadSize;
93     // The offset of this load from the base pointer, in bytes.
94     uint64_t Offset;
95   };
96   using LoadEntryVector = SmallVector<LoadEntry, 8>;
97   LoadEntryVector LoadSequence;
98 
99   void createLoadCmpBlocks();
100   void createResultBlock();
101   void setupResultBlockPHINodes();
102   void setupEndBlockPHINodes();
103   Value *getCompareLoadPairs(unsigned BlockIndex, unsigned &LoadIndex);
104   void emitLoadCompareBlock(unsigned BlockIndex);
105   void emitLoadCompareBlockMultipleLoads(unsigned BlockIndex,
106                                          unsigned &LoadIndex);
107   void emitLoadCompareByteBlock(unsigned BlockIndex, unsigned OffsetBytes);
108   void emitMemCmpResultBlock();
109   Value *getMemCmpExpansionZeroCase();
110   Value *getMemCmpEqZeroOneBlock();
111   Value *getMemCmpOneBlock();
112   struct LoadPair {
113     Value *Lhs = nullptr;
114     Value *Rhs = nullptr;
115   };
116   LoadPair getLoadPair(Type *LoadSizeType, bool NeedsBSwap, Type *CmpSizeType,
117                        unsigned OffsetBytes);
118 
119   static LoadEntryVector
120   computeGreedyLoadSequence(uint64_t Size, llvm::ArrayRef<unsigned> LoadSizes,
121                             unsigned MaxNumLoads, unsigned &NumLoadsNonOneByte);
122   static LoadEntryVector
123   computeOverlappingLoadSequence(uint64_t Size, unsigned MaxLoadSize,
124                                  unsigned MaxNumLoads,
125                                  unsigned &NumLoadsNonOneByte);
126 
127 public:
128   MemCmpExpansion(CallInst *CI, uint64_t Size,
129                   const TargetTransformInfo::MemCmpExpansionOptions &Options,
130                   const bool IsUsedForZeroCmp, const DataLayout &TheDataLayout,
131                   DomTreeUpdater *DTU);
132 
133   unsigned getNumBlocks();
getNumLoads() const134   uint64_t getNumLoads() const { return LoadSequence.size(); }
135 
136   Value *getMemCmpExpansion();
137 };
138 
computeGreedyLoadSequence(uint64_t Size,llvm::ArrayRef<unsigned> LoadSizes,const unsigned MaxNumLoads,unsigned & NumLoadsNonOneByte)139 MemCmpExpansion::LoadEntryVector MemCmpExpansion::computeGreedyLoadSequence(
140     uint64_t Size, llvm::ArrayRef<unsigned> LoadSizes,
141     const unsigned MaxNumLoads, unsigned &NumLoadsNonOneByte) {
142   NumLoadsNonOneByte = 0;
143   LoadEntryVector LoadSequence;
144   uint64_t Offset = 0;
145   while (Size && !LoadSizes.empty()) {
146     const unsigned LoadSize = LoadSizes.front();
147     const uint64_t NumLoadsForThisSize = Size / LoadSize;
148     if (LoadSequence.size() + NumLoadsForThisSize > MaxNumLoads) {
149       // Do not expand if the total number of loads is larger than what the
150       // target allows. Note that it's important that we exit before completing
151       // the expansion to avoid using a ton of memory to store the expansion for
152       // large sizes.
153       return {};
154     }
155     if (NumLoadsForThisSize > 0) {
156       for (uint64_t I = 0; I < NumLoadsForThisSize; ++I) {
157         LoadSequence.push_back({LoadSize, Offset});
158         Offset += LoadSize;
159       }
160       if (LoadSize > 1)
161         ++NumLoadsNonOneByte;
162       Size = Size % LoadSize;
163     }
164     LoadSizes = LoadSizes.drop_front();
165   }
166   return LoadSequence;
167 }
168 
169 MemCmpExpansion::LoadEntryVector
computeOverlappingLoadSequence(uint64_t Size,const unsigned MaxLoadSize,const unsigned MaxNumLoads,unsigned & NumLoadsNonOneByte)170 MemCmpExpansion::computeOverlappingLoadSequence(uint64_t Size,
171                                                 const unsigned MaxLoadSize,
172                                                 const unsigned MaxNumLoads,
173                                                 unsigned &NumLoadsNonOneByte) {
174   // These are already handled by the greedy approach.
175   if (Size < 2 || MaxLoadSize < 2)
176     return {};
177 
178   // We try to do as many non-overlapping loads as possible starting from the
179   // beginning.
180   const uint64_t NumNonOverlappingLoads = Size / MaxLoadSize;
181   assert(NumNonOverlappingLoads && "there must be at least one load");
182   // There remain 0 to (MaxLoadSize - 1) bytes to load, this will be done with
183   // an overlapping load.
184   Size = Size - NumNonOverlappingLoads * MaxLoadSize;
185   // Bail if we do not need an overloapping store, this is already handled by
186   // the greedy approach.
187   if (Size == 0)
188     return {};
189   // Bail if the number of loads (non-overlapping + potential overlapping one)
190   // is larger than the max allowed.
191   if ((NumNonOverlappingLoads + 1) > MaxNumLoads)
192     return {};
193 
194   // Add non-overlapping loads.
195   LoadEntryVector LoadSequence;
196   uint64_t Offset = 0;
197   for (uint64_t I = 0; I < NumNonOverlappingLoads; ++I) {
198     LoadSequence.push_back({MaxLoadSize, Offset});
199     Offset += MaxLoadSize;
200   }
201 
202   // Add the last overlapping load.
203   assert(Size > 0 && Size < MaxLoadSize && "broken invariant");
204   LoadSequence.push_back({MaxLoadSize, Offset - (MaxLoadSize - Size)});
205   NumLoadsNonOneByte = 1;
206   return LoadSequence;
207 }
208 
209 // Initialize the basic block structure required for expansion of memcmp call
210 // with given maximum load size and memcmp size parameter.
211 // This structure includes:
212 // 1. A list of load compare blocks - LoadCmpBlocks.
213 // 2. An EndBlock, split from original instruction point, which is the block to
214 // return from.
215 // 3. ResultBlock, block to branch to for early exit when a
216 // LoadCmpBlock finds a difference.
MemCmpExpansion(CallInst * const CI,uint64_t Size,const TargetTransformInfo::MemCmpExpansionOptions & Options,const bool IsUsedForZeroCmp,const DataLayout & TheDataLayout,DomTreeUpdater * DTU)217 MemCmpExpansion::MemCmpExpansion(
218     CallInst *const CI, uint64_t Size,
219     const TargetTransformInfo::MemCmpExpansionOptions &Options,
220     const bool IsUsedForZeroCmp, const DataLayout &TheDataLayout,
221     DomTreeUpdater *DTU)
222     : CI(CI), Size(Size), MaxLoadSize(0), NumLoadsNonOneByte(0),
223       NumLoadsPerBlockForZeroCmp(Options.NumLoadsPerBlock),
224       IsUsedForZeroCmp(IsUsedForZeroCmp), DL(TheDataLayout), DTU(DTU),
225       Builder(CI) {
226   assert(Size > 0 && "zero blocks");
227   // Scale the max size down if the target can load more bytes than we need.
228   llvm::ArrayRef<unsigned> LoadSizes(Options.LoadSizes);
229   while (!LoadSizes.empty() && LoadSizes.front() > Size) {
230     LoadSizes = LoadSizes.drop_front();
231   }
232   assert(!LoadSizes.empty() && "cannot load Size bytes");
233   MaxLoadSize = LoadSizes.front();
234   // Compute the decomposition.
235   unsigned GreedyNumLoadsNonOneByte = 0;
236   LoadSequence = computeGreedyLoadSequence(Size, LoadSizes, Options.MaxNumLoads,
237                                            GreedyNumLoadsNonOneByte);
238   NumLoadsNonOneByte = GreedyNumLoadsNonOneByte;
239   assert(LoadSequence.size() <= Options.MaxNumLoads && "broken invariant");
240   // If we allow overlapping loads and the load sequence is not already optimal,
241   // use overlapping loads.
242   if (Options.AllowOverlappingLoads &&
243       (LoadSequence.empty() || LoadSequence.size() > 2)) {
244     unsigned OverlappingNumLoadsNonOneByte = 0;
245     auto OverlappingLoads = computeOverlappingLoadSequence(
246         Size, MaxLoadSize, Options.MaxNumLoads, OverlappingNumLoadsNonOneByte);
247     if (!OverlappingLoads.empty() &&
248         (LoadSequence.empty() ||
249          OverlappingLoads.size() < LoadSequence.size())) {
250       LoadSequence = OverlappingLoads;
251       NumLoadsNonOneByte = OverlappingNumLoadsNonOneByte;
252     }
253   }
254   assert(LoadSequence.size() <= Options.MaxNumLoads && "broken invariant");
255 }
256 
getNumBlocks()257 unsigned MemCmpExpansion::getNumBlocks() {
258   if (IsUsedForZeroCmp)
259     return getNumLoads() / NumLoadsPerBlockForZeroCmp +
260            (getNumLoads() % NumLoadsPerBlockForZeroCmp != 0 ? 1 : 0);
261   return getNumLoads();
262 }
263 
createLoadCmpBlocks()264 void MemCmpExpansion::createLoadCmpBlocks() {
265   for (unsigned i = 0; i < getNumBlocks(); i++) {
266     BasicBlock *BB = BasicBlock::Create(CI->getContext(), "loadbb",
267                                         EndBlock->getParent(), EndBlock);
268     LoadCmpBlocks.push_back(BB);
269   }
270 }
271 
createResultBlock()272 void MemCmpExpansion::createResultBlock() {
273   ResBlock.BB = BasicBlock::Create(CI->getContext(), "res_block",
274                                    EndBlock->getParent(), EndBlock);
275 }
276 
getLoadPair(Type * LoadSizeType,bool NeedsBSwap,Type * CmpSizeType,unsigned OffsetBytes)277 MemCmpExpansion::LoadPair MemCmpExpansion::getLoadPair(Type *LoadSizeType,
278                                                        bool NeedsBSwap,
279                                                        Type *CmpSizeType,
280                                                        unsigned OffsetBytes) {
281   // Get the memory source at offset `OffsetBytes`.
282   Value *LhsSource = CI->getArgOperand(0);
283   Value *RhsSource = CI->getArgOperand(1);
284   Align LhsAlign = LhsSource->getPointerAlignment(DL);
285   Align RhsAlign = RhsSource->getPointerAlignment(DL);
286   if (OffsetBytes > 0) {
287     auto *ByteType = Type::getInt8Ty(CI->getContext());
288     LhsSource = Builder.CreateConstGEP1_64(
289         ByteType, Builder.CreateBitCast(LhsSource, ByteType->getPointerTo()),
290         OffsetBytes);
291     RhsSource = Builder.CreateConstGEP1_64(
292         ByteType, Builder.CreateBitCast(RhsSource, ByteType->getPointerTo()),
293         OffsetBytes);
294     LhsAlign = commonAlignment(LhsAlign, OffsetBytes);
295     RhsAlign = commonAlignment(RhsAlign, OffsetBytes);
296   }
297   LhsSource = Builder.CreateBitCast(LhsSource, LoadSizeType->getPointerTo());
298   RhsSource = Builder.CreateBitCast(RhsSource, LoadSizeType->getPointerTo());
299 
300   // Create a constant or a load from the source.
301   Value *Lhs = nullptr;
302   if (auto *C = dyn_cast<Constant>(LhsSource))
303     Lhs = ConstantFoldLoadFromConstPtr(C, LoadSizeType, DL);
304   if (!Lhs)
305     Lhs = Builder.CreateAlignedLoad(LoadSizeType, LhsSource, LhsAlign);
306 
307   Value *Rhs = nullptr;
308   if (auto *C = dyn_cast<Constant>(RhsSource))
309     Rhs = ConstantFoldLoadFromConstPtr(C, LoadSizeType, DL);
310   if (!Rhs)
311     Rhs = Builder.CreateAlignedLoad(LoadSizeType, RhsSource, RhsAlign);
312 
313   // Swap bytes if required.
314   if (NeedsBSwap) {
315     Function *Bswap = Intrinsic::getDeclaration(CI->getModule(),
316                                                 Intrinsic::bswap, LoadSizeType);
317     Lhs = Builder.CreateCall(Bswap, Lhs);
318     Rhs = Builder.CreateCall(Bswap, Rhs);
319   }
320 
321   // Zero extend if required.
322   if (CmpSizeType != nullptr && CmpSizeType != LoadSizeType) {
323     Lhs = Builder.CreateZExt(Lhs, CmpSizeType);
324     Rhs = Builder.CreateZExt(Rhs, CmpSizeType);
325   }
326   return {Lhs, Rhs};
327 }
328 
329 // This function creates the IR instructions for loading and comparing 1 byte.
330 // It loads 1 byte from each source of the memcmp parameters with the given
331 // GEPIndex. It then subtracts the two loaded values and adds this result to the
332 // final phi node for selecting the memcmp result.
emitLoadCompareByteBlock(unsigned BlockIndex,unsigned OffsetBytes)333 void MemCmpExpansion::emitLoadCompareByteBlock(unsigned BlockIndex,
334                                                unsigned OffsetBytes) {
335   BasicBlock *BB = LoadCmpBlocks[BlockIndex];
336   Builder.SetInsertPoint(BB);
337   const LoadPair Loads =
338       getLoadPair(Type::getInt8Ty(CI->getContext()), /*NeedsBSwap=*/false,
339                   Type::getInt32Ty(CI->getContext()), OffsetBytes);
340   Value *Diff = Builder.CreateSub(Loads.Lhs, Loads.Rhs);
341 
342   PhiRes->addIncoming(Diff, BB);
343 
344   if (BlockIndex < (LoadCmpBlocks.size() - 1)) {
345     // Early exit branch if difference found to EndBlock. Otherwise, continue to
346     // next LoadCmpBlock,
347     Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_NE, Diff,
348                                     ConstantInt::get(Diff->getType(), 0));
349     BranchInst *CmpBr =
350         BranchInst::Create(EndBlock, LoadCmpBlocks[BlockIndex + 1], Cmp);
351     if (DTU)
352       DTU->applyUpdates(
353           {{DominatorTree::Insert, BB, EndBlock},
354            {DominatorTree::Insert, BB, LoadCmpBlocks[BlockIndex + 1]}});
355     Builder.Insert(CmpBr);
356   } else {
357     // The last block has an unconditional branch to EndBlock.
358     BranchInst *CmpBr = BranchInst::Create(EndBlock);
359     if (DTU)
360       DTU->applyUpdates({{DominatorTree::Insert, BB, EndBlock}});
361     Builder.Insert(CmpBr);
362   }
363 }
364 
365 /// Generate an equality comparison for one or more pairs of loaded values.
366 /// This is used in the case where the memcmp() call is compared equal or not
367 /// equal to zero.
getCompareLoadPairs(unsigned BlockIndex,unsigned & LoadIndex)368 Value *MemCmpExpansion::getCompareLoadPairs(unsigned BlockIndex,
369                                             unsigned &LoadIndex) {
370   assert(LoadIndex < getNumLoads() &&
371          "getCompareLoadPairs() called with no remaining loads");
372   std::vector<Value *> XorList, OrList;
373   Value *Diff = nullptr;
374 
375   const unsigned NumLoads =
376       std::min(getNumLoads() - LoadIndex, NumLoadsPerBlockForZeroCmp);
377 
378   // For a single-block expansion, start inserting before the memcmp call.
379   if (LoadCmpBlocks.empty())
380     Builder.SetInsertPoint(CI);
381   else
382     Builder.SetInsertPoint(LoadCmpBlocks[BlockIndex]);
383 
384   Value *Cmp = nullptr;
385   // If we have multiple loads per block, we need to generate a composite
386   // comparison using xor+or. The type for the combinations is the largest load
387   // type.
388   IntegerType *const MaxLoadType =
389       NumLoads == 1 ? nullptr
390                     : IntegerType::get(CI->getContext(), MaxLoadSize * 8);
391   for (unsigned i = 0; i < NumLoads; ++i, ++LoadIndex) {
392     const LoadEntry &CurLoadEntry = LoadSequence[LoadIndex];
393     const LoadPair Loads = getLoadPair(
394         IntegerType::get(CI->getContext(), CurLoadEntry.LoadSize * 8),
395         /*NeedsBSwap=*/false, MaxLoadType, CurLoadEntry.Offset);
396 
397     if (NumLoads != 1) {
398       // If we have multiple loads per block, we need to generate a composite
399       // comparison using xor+or.
400       Diff = Builder.CreateXor(Loads.Lhs, Loads.Rhs);
401       Diff = Builder.CreateZExt(Diff, MaxLoadType);
402       XorList.push_back(Diff);
403     } else {
404       // If there's only one load per block, we just compare the loaded values.
405       Cmp = Builder.CreateICmpNE(Loads.Lhs, Loads.Rhs);
406     }
407   }
408 
409   auto pairWiseOr = [&](std::vector<Value *> &InList) -> std::vector<Value *> {
410     std::vector<Value *> OutList;
411     for (unsigned i = 0; i < InList.size() - 1; i = i + 2) {
412       Value *Or = Builder.CreateOr(InList[i], InList[i + 1]);
413       OutList.push_back(Or);
414     }
415     if (InList.size() % 2 != 0)
416       OutList.push_back(InList.back());
417     return OutList;
418   };
419 
420   if (!Cmp) {
421     // Pairwise OR the XOR results.
422     OrList = pairWiseOr(XorList);
423 
424     // Pairwise OR the OR results until one result left.
425     while (OrList.size() != 1) {
426       OrList = pairWiseOr(OrList);
427     }
428 
429     assert(Diff && "Failed to find comparison diff");
430     Cmp = Builder.CreateICmpNE(OrList[0], ConstantInt::get(Diff->getType(), 0));
431   }
432 
433   return Cmp;
434 }
435 
emitLoadCompareBlockMultipleLoads(unsigned BlockIndex,unsigned & LoadIndex)436 void MemCmpExpansion::emitLoadCompareBlockMultipleLoads(unsigned BlockIndex,
437                                                         unsigned &LoadIndex) {
438   Value *Cmp = getCompareLoadPairs(BlockIndex, LoadIndex);
439 
440   BasicBlock *NextBB = (BlockIndex == (LoadCmpBlocks.size() - 1))
441                            ? EndBlock
442                            : LoadCmpBlocks[BlockIndex + 1];
443   // Early exit branch if difference found to ResultBlock. Otherwise,
444   // continue to next LoadCmpBlock or EndBlock.
445   BasicBlock *BB = Builder.GetInsertBlock();
446   BranchInst *CmpBr = BranchInst::Create(ResBlock.BB, NextBB, Cmp);
447   Builder.Insert(CmpBr);
448   if (DTU)
449     DTU->applyUpdates({{DominatorTree::Insert, BB, ResBlock.BB},
450                        {DominatorTree::Insert, BB, NextBB}});
451 
452   // Add a phi edge for the last LoadCmpBlock to Endblock with a value of 0
453   // since early exit to ResultBlock was not taken (no difference was found in
454   // any of the bytes).
455   if (BlockIndex == LoadCmpBlocks.size() - 1) {
456     Value *Zero = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 0);
457     PhiRes->addIncoming(Zero, LoadCmpBlocks[BlockIndex]);
458   }
459 }
460 
461 // This function creates the IR intructions for loading and comparing using the
462 // given LoadSize. It loads the number of bytes specified by LoadSize from each
463 // source of the memcmp parameters. It then does a subtract to see if there was
464 // a difference in the loaded values. If a difference is found, it branches
465 // with an early exit to the ResultBlock for calculating which source was
466 // larger. Otherwise, it falls through to the either the next LoadCmpBlock or
467 // the EndBlock if this is the last LoadCmpBlock. Loading 1 byte is handled with
468 // a special case through emitLoadCompareByteBlock. The special handling can
469 // simply subtract the loaded values and add it to the result phi node.
emitLoadCompareBlock(unsigned BlockIndex)470 void MemCmpExpansion::emitLoadCompareBlock(unsigned BlockIndex) {
471   // There is one load per block in this case, BlockIndex == LoadIndex.
472   const LoadEntry &CurLoadEntry = LoadSequence[BlockIndex];
473 
474   if (CurLoadEntry.LoadSize == 1) {
475     MemCmpExpansion::emitLoadCompareByteBlock(BlockIndex, CurLoadEntry.Offset);
476     return;
477   }
478 
479   Type *LoadSizeType =
480       IntegerType::get(CI->getContext(), CurLoadEntry.LoadSize * 8);
481   Type *MaxLoadType = IntegerType::get(CI->getContext(), MaxLoadSize * 8);
482   assert(CurLoadEntry.LoadSize <= MaxLoadSize && "Unexpected load type");
483 
484   Builder.SetInsertPoint(LoadCmpBlocks[BlockIndex]);
485 
486   const LoadPair Loads =
487       getLoadPair(LoadSizeType, /*NeedsBSwap=*/DL.isLittleEndian(), MaxLoadType,
488                   CurLoadEntry.Offset);
489 
490   // Add the loaded values to the phi nodes for calculating memcmp result only
491   // if result is not used in a zero equality.
492   if (!IsUsedForZeroCmp) {
493     ResBlock.PhiSrc1->addIncoming(Loads.Lhs, LoadCmpBlocks[BlockIndex]);
494     ResBlock.PhiSrc2->addIncoming(Loads.Rhs, LoadCmpBlocks[BlockIndex]);
495   }
496 
497   Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Loads.Lhs, Loads.Rhs);
498   BasicBlock *NextBB = (BlockIndex == (LoadCmpBlocks.size() - 1))
499                            ? EndBlock
500                            : LoadCmpBlocks[BlockIndex + 1];
501   // Early exit branch if difference found to ResultBlock. Otherwise, continue
502   // to next LoadCmpBlock or EndBlock.
503   BasicBlock *BB = Builder.GetInsertBlock();
504   BranchInst *CmpBr = BranchInst::Create(NextBB, ResBlock.BB, Cmp);
505   Builder.Insert(CmpBr);
506   if (DTU)
507     DTU->applyUpdates({{DominatorTree::Insert, BB, NextBB},
508                        {DominatorTree::Insert, BB, ResBlock.BB}});
509 
510   // Add a phi edge for the last LoadCmpBlock to Endblock with a value of 0
511   // since early exit to ResultBlock was not taken (no difference was found in
512   // any of the bytes).
513   if (BlockIndex == LoadCmpBlocks.size() - 1) {
514     Value *Zero = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 0);
515     PhiRes->addIncoming(Zero, LoadCmpBlocks[BlockIndex]);
516   }
517 }
518 
519 // This function populates the ResultBlock with a sequence to calculate the
520 // memcmp result. It compares the two loaded source values and returns -1 if
521 // src1 < src2 and 1 if src1 > src2.
emitMemCmpResultBlock()522 void MemCmpExpansion::emitMemCmpResultBlock() {
523   // Special case: if memcmp result is used in a zero equality, result does not
524   // need to be calculated and can simply return 1.
525   if (IsUsedForZeroCmp) {
526     BasicBlock::iterator InsertPt = ResBlock.BB->getFirstInsertionPt();
527     Builder.SetInsertPoint(ResBlock.BB, InsertPt);
528     Value *Res = ConstantInt::get(Type::getInt32Ty(CI->getContext()), 1);
529     PhiRes->addIncoming(Res, ResBlock.BB);
530     BranchInst *NewBr = BranchInst::Create(EndBlock);
531     Builder.Insert(NewBr);
532     if (DTU)
533       DTU->applyUpdates({{DominatorTree::Insert, ResBlock.BB, EndBlock}});
534     return;
535   }
536   BasicBlock::iterator InsertPt = ResBlock.BB->getFirstInsertionPt();
537   Builder.SetInsertPoint(ResBlock.BB, InsertPt);
538 
539   Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_ULT, ResBlock.PhiSrc1,
540                                   ResBlock.PhiSrc2);
541 
542   Value *Res =
543       Builder.CreateSelect(Cmp, ConstantInt::get(Builder.getInt32Ty(), -1),
544                            ConstantInt::get(Builder.getInt32Ty(), 1));
545 
546   PhiRes->addIncoming(Res, ResBlock.BB);
547   BranchInst *NewBr = BranchInst::Create(EndBlock);
548   Builder.Insert(NewBr);
549   if (DTU)
550     DTU->applyUpdates({{DominatorTree::Insert, ResBlock.BB, EndBlock}});
551 }
552 
setupResultBlockPHINodes()553 void MemCmpExpansion::setupResultBlockPHINodes() {
554   Type *MaxLoadType = IntegerType::get(CI->getContext(), MaxLoadSize * 8);
555   Builder.SetInsertPoint(ResBlock.BB);
556   // Note: this assumes one load per block.
557   ResBlock.PhiSrc1 =
558       Builder.CreatePHI(MaxLoadType, NumLoadsNonOneByte, "phi.src1");
559   ResBlock.PhiSrc2 =
560       Builder.CreatePHI(MaxLoadType, NumLoadsNonOneByte, "phi.src2");
561 }
562 
setupEndBlockPHINodes()563 void MemCmpExpansion::setupEndBlockPHINodes() {
564   Builder.SetInsertPoint(&EndBlock->front());
565   PhiRes = Builder.CreatePHI(Type::getInt32Ty(CI->getContext()), 2, "phi.res");
566 }
567 
getMemCmpExpansionZeroCase()568 Value *MemCmpExpansion::getMemCmpExpansionZeroCase() {
569   unsigned LoadIndex = 0;
570   // This loop populates each of the LoadCmpBlocks with the IR sequence to
571   // handle multiple loads per block.
572   for (unsigned I = 0; I < getNumBlocks(); ++I) {
573     emitLoadCompareBlockMultipleLoads(I, LoadIndex);
574   }
575 
576   emitMemCmpResultBlock();
577   return PhiRes;
578 }
579 
580 /// A memcmp expansion that compares equality with 0 and only has one block of
581 /// load and compare can bypass the compare, branch, and phi IR that is required
582 /// in the general case.
getMemCmpEqZeroOneBlock()583 Value *MemCmpExpansion::getMemCmpEqZeroOneBlock() {
584   unsigned LoadIndex = 0;
585   Value *Cmp = getCompareLoadPairs(0, LoadIndex);
586   assert(LoadIndex == getNumLoads() && "some entries were not consumed");
587   return Builder.CreateZExt(Cmp, Type::getInt32Ty(CI->getContext()));
588 }
589 
590 /// A memcmp expansion that only has one block of load and compare can bypass
591 /// the compare, branch, and phi IR that is required in the general case.
getMemCmpOneBlock()592 Value *MemCmpExpansion::getMemCmpOneBlock() {
593   Type *LoadSizeType = IntegerType::get(CI->getContext(), Size * 8);
594   bool NeedsBSwap = DL.isLittleEndian() && Size != 1;
595 
596   // The i8 and i16 cases don't need compares. We zext the loaded values and
597   // subtract them to get the suitable negative, zero, or positive i32 result.
598   if (Size < 4) {
599     const LoadPair Loads =
600         getLoadPair(LoadSizeType, NeedsBSwap, Builder.getInt32Ty(),
601                     /*Offset*/ 0);
602     return Builder.CreateSub(Loads.Lhs, Loads.Rhs);
603   }
604 
605   const LoadPair Loads = getLoadPair(LoadSizeType, NeedsBSwap, LoadSizeType,
606                                      /*Offset*/ 0);
607   // The result of memcmp is negative, zero, or positive, so produce that by
608   // subtracting 2 extended compare bits: sub (ugt, ult).
609   // If a target prefers to use selects to get -1/0/1, they should be able
610   // to transform this later. The inverse transform (going from selects to math)
611   // may not be possible in the DAG because the selects got converted into
612   // branches before we got there.
613   Value *CmpUGT = Builder.CreateICmpUGT(Loads.Lhs, Loads.Rhs);
614   Value *CmpULT = Builder.CreateICmpULT(Loads.Lhs, Loads.Rhs);
615   Value *ZextUGT = Builder.CreateZExt(CmpUGT, Builder.getInt32Ty());
616   Value *ZextULT = Builder.CreateZExt(CmpULT, Builder.getInt32Ty());
617   return Builder.CreateSub(ZextUGT, ZextULT);
618 }
619 
620 // This function expands the memcmp call into an inline expansion and returns
621 // the memcmp result.
getMemCmpExpansion()622 Value *MemCmpExpansion::getMemCmpExpansion() {
623   // Create the basic block framework for a multi-block expansion.
624   if (getNumBlocks() != 1) {
625     BasicBlock *StartBlock = CI->getParent();
626     EndBlock = SplitBlock(StartBlock, CI, DTU, /*LI=*/nullptr,
627                           /*MSSAU=*/nullptr, "endblock");
628     setupEndBlockPHINodes();
629     createResultBlock();
630 
631     // If return value of memcmp is not used in a zero equality, we need to
632     // calculate which source was larger. The calculation requires the
633     // two loaded source values of each load compare block.
634     // These will be saved in the phi nodes created by setupResultBlockPHINodes.
635     if (!IsUsedForZeroCmp) setupResultBlockPHINodes();
636 
637     // Create the number of required load compare basic blocks.
638     createLoadCmpBlocks();
639 
640     // Update the terminator added by SplitBlock to branch to the first
641     // LoadCmpBlock.
642     StartBlock->getTerminator()->setSuccessor(0, LoadCmpBlocks[0]);
643     if (DTU)
644       DTU->applyUpdates({{DominatorTree::Insert, StartBlock, LoadCmpBlocks[0]},
645                          {DominatorTree::Delete, StartBlock, EndBlock}});
646   }
647 
648   Builder.SetCurrentDebugLocation(CI->getDebugLoc());
649 
650   if (IsUsedForZeroCmp)
651     return getNumBlocks() == 1 ? getMemCmpEqZeroOneBlock()
652                                : getMemCmpExpansionZeroCase();
653 
654   if (getNumBlocks() == 1)
655     return getMemCmpOneBlock();
656 
657   for (unsigned I = 0; I < getNumBlocks(); ++I) {
658     emitLoadCompareBlock(I);
659   }
660 
661   emitMemCmpResultBlock();
662   return PhiRes;
663 }
664 
665 // This function checks to see if an expansion of memcmp can be generated.
666 // It checks for constant compare size that is less than the max inline size.
667 // If an expansion cannot occur, returns false to leave as a library call.
668 // Otherwise, the library call is replaced with a new IR instruction sequence.
669 /// We want to transform:
670 /// %call = call signext i32 @memcmp(i8* %0, i8* %1, i64 15)
671 /// To:
672 /// loadbb:
673 ///  %0 = bitcast i32* %buffer2 to i8*
674 ///  %1 = bitcast i32* %buffer1 to i8*
675 ///  %2 = bitcast i8* %1 to i64*
676 ///  %3 = bitcast i8* %0 to i64*
677 ///  %4 = load i64, i64* %2
678 ///  %5 = load i64, i64* %3
679 ///  %6 = call i64 @llvm.bswap.i64(i64 %4)
680 ///  %7 = call i64 @llvm.bswap.i64(i64 %5)
681 ///  %8 = sub i64 %6, %7
682 ///  %9 = icmp ne i64 %8, 0
683 ///  br i1 %9, label %res_block, label %loadbb1
684 /// res_block:                                        ; preds = %loadbb2,
685 /// %loadbb1, %loadbb
686 ///  %phi.src1 = phi i64 [ %6, %loadbb ], [ %22, %loadbb1 ], [ %36, %loadbb2 ]
687 ///  %phi.src2 = phi i64 [ %7, %loadbb ], [ %23, %loadbb1 ], [ %37, %loadbb2 ]
688 ///  %10 = icmp ult i64 %phi.src1, %phi.src2
689 ///  %11 = select i1 %10, i32 -1, i32 1
690 ///  br label %endblock
691 /// loadbb1:                                          ; preds = %loadbb
692 ///  %12 = bitcast i32* %buffer2 to i8*
693 ///  %13 = bitcast i32* %buffer1 to i8*
694 ///  %14 = bitcast i8* %13 to i32*
695 ///  %15 = bitcast i8* %12 to i32*
696 ///  %16 = getelementptr i32, i32* %14, i32 2
697 ///  %17 = getelementptr i32, i32* %15, i32 2
698 ///  %18 = load i32, i32* %16
699 ///  %19 = load i32, i32* %17
700 ///  %20 = call i32 @llvm.bswap.i32(i32 %18)
701 ///  %21 = call i32 @llvm.bswap.i32(i32 %19)
702 ///  %22 = zext i32 %20 to i64
703 ///  %23 = zext i32 %21 to i64
704 ///  %24 = sub i64 %22, %23
705 ///  %25 = icmp ne i64 %24, 0
706 ///  br i1 %25, label %res_block, label %loadbb2
707 /// loadbb2:                                          ; preds = %loadbb1
708 ///  %26 = bitcast i32* %buffer2 to i8*
709 ///  %27 = bitcast i32* %buffer1 to i8*
710 ///  %28 = bitcast i8* %27 to i16*
711 ///  %29 = bitcast i8* %26 to i16*
712 ///  %30 = getelementptr i16, i16* %28, i16 6
713 ///  %31 = getelementptr i16, i16* %29, i16 6
714 ///  %32 = load i16, i16* %30
715 ///  %33 = load i16, i16* %31
716 ///  %34 = call i16 @llvm.bswap.i16(i16 %32)
717 ///  %35 = call i16 @llvm.bswap.i16(i16 %33)
718 ///  %36 = zext i16 %34 to i64
719 ///  %37 = zext i16 %35 to i64
720 ///  %38 = sub i64 %36, %37
721 ///  %39 = icmp ne i64 %38, 0
722 ///  br i1 %39, label %res_block, label %loadbb3
723 /// loadbb3:                                          ; preds = %loadbb2
724 ///  %40 = bitcast i32* %buffer2 to i8*
725 ///  %41 = bitcast i32* %buffer1 to i8*
726 ///  %42 = getelementptr i8, i8* %41, i8 14
727 ///  %43 = getelementptr i8, i8* %40, i8 14
728 ///  %44 = load i8, i8* %42
729 ///  %45 = load i8, i8* %43
730 ///  %46 = zext i8 %44 to i32
731 ///  %47 = zext i8 %45 to i32
732 ///  %48 = sub i32 %46, %47
733 ///  br label %endblock
734 /// endblock:                                         ; preds = %res_block,
735 /// %loadbb3
736 ///  %phi.res = phi i32 [ %48, %loadbb3 ], [ %11, %res_block ]
737 ///  ret i32 %phi.res
expandMemCmp(CallInst * CI,const TargetTransformInfo * TTI,const TargetLowering * TLI,const DataLayout * DL,ProfileSummaryInfo * PSI,BlockFrequencyInfo * BFI,DomTreeUpdater * DTU)738 static bool expandMemCmp(CallInst *CI, const TargetTransformInfo *TTI,
739                          const TargetLowering *TLI, const DataLayout *DL,
740                          ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI,
741                          DomTreeUpdater *DTU) {
742   NumMemCmpCalls++;
743 
744   // Early exit from expansion if -Oz.
745   if (CI->getFunction()->hasMinSize())
746     return false;
747 
748   // Early exit from expansion if size is not a constant.
749   ConstantInt *SizeCast = dyn_cast<ConstantInt>(CI->getArgOperand(2));
750   if (!SizeCast) {
751     NumMemCmpNotConstant++;
752     return false;
753   }
754   const uint64_t SizeVal = SizeCast->getZExtValue();
755 
756   if (SizeVal == 0) {
757     return false;
758   }
759   // TTI call to check if target would like to expand memcmp. Also, get the
760   // available load sizes.
761   const bool IsUsedForZeroCmp = isOnlyUsedInZeroEqualityComparison(CI);
762   bool OptForSize = CI->getFunction()->hasOptSize() ||
763                     llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI);
764   auto Options = TTI->enableMemCmpExpansion(OptForSize,
765                                             IsUsedForZeroCmp);
766   if (!Options) return false;
767 
768   if (MemCmpEqZeroNumLoadsPerBlock.getNumOccurrences())
769     Options.NumLoadsPerBlock = MemCmpEqZeroNumLoadsPerBlock;
770 
771   if (OptForSize &&
772       MaxLoadsPerMemcmpOptSize.getNumOccurrences())
773     Options.MaxNumLoads = MaxLoadsPerMemcmpOptSize;
774 
775   if (!OptForSize && MaxLoadsPerMemcmp.getNumOccurrences())
776     Options.MaxNumLoads = MaxLoadsPerMemcmp;
777 
778   MemCmpExpansion Expansion(CI, SizeVal, Options, IsUsedForZeroCmp, *DL, DTU);
779 
780   // Don't expand if this will require more loads than desired by the target.
781   if (Expansion.getNumLoads() == 0) {
782     NumMemCmpGreaterThanMax++;
783     return false;
784   }
785 
786   NumMemCmpInlined++;
787 
788   Value *Res = Expansion.getMemCmpExpansion();
789 
790   // Replace call with result of expansion and erase call.
791   CI->replaceAllUsesWith(Res);
792   CI->eraseFromParent();
793 
794   return true;
795 }
796 
797 class ExpandMemCmpPass : public FunctionPass {
798 public:
799   static char ID;
800 
ExpandMemCmpPass()801   ExpandMemCmpPass() : FunctionPass(ID) {
802     initializeExpandMemCmpPassPass(*PassRegistry::getPassRegistry());
803   }
804 
runOnFunction(Function & F)805   bool runOnFunction(Function &F) override {
806     if (skipFunction(F)) return false;
807 
808     auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
809     if (!TPC) {
810       return false;
811     }
812     const TargetLowering* TL =
813         TPC->getTM<TargetMachine>().getSubtargetImpl(F)->getTargetLowering();
814 
815     const TargetLibraryInfo *TLI =
816         &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
817     const TargetTransformInfo *TTI =
818         &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
819     auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
820     auto *BFI = (PSI && PSI->hasProfileSummary()) ?
821            &getAnalysis<LazyBlockFrequencyInfoPass>().getBFI() :
822            nullptr;
823     DominatorTree *DT = nullptr;
824     if (auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>())
825       DT = &DTWP->getDomTree();
826     auto PA = runImpl(F, TLI, TTI, TL, PSI, BFI, DT);
827     return !PA.areAllPreserved();
828   }
829 
830 private:
getAnalysisUsage(AnalysisUsage & AU) const831   void getAnalysisUsage(AnalysisUsage &AU) const override {
832     AU.addRequired<TargetLibraryInfoWrapperPass>();
833     AU.addRequired<TargetTransformInfoWrapperPass>();
834     AU.addRequired<ProfileSummaryInfoWrapperPass>();
835     AU.addPreserved<DominatorTreeWrapperPass>();
836     LazyBlockFrequencyInfoPass::getLazyBFIAnalysisUsage(AU);
837     FunctionPass::getAnalysisUsage(AU);
838   }
839 
840   PreservedAnalyses runImpl(Function &F, const TargetLibraryInfo *TLI,
841                             const TargetTransformInfo *TTI,
842                             const TargetLowering *TL, ProfileSummaryInfo *PSI,
843                             BlockFrequencyInfo *BFI, DominatorTree *DT);
844   // Returns true if a change was made.
845   bool runOnBlock(BasicBlock &BB, const TargetLibraryInfo *TLI,
846                   const TargetTransformInfo *TTI, const TargetLowering *TL,
847                   const DataLayout &DL, ProfileSummaryInfo *PSI,
848                   BlockFrequencyInfo *BFI, DomTreeUpdater *DTU);
849 };
850 
runOnBlock(BasicBlock & BB,const TargetLibraryInfo * TLI,const TargetTransformInfo * TTI,const TargetLowering * TL,const DataLayout & DL,ProfileSummaryInfo * PSI,BlockFrequencyInfo * BFI,DomTreeUpdater * DTU)851 bool ExpandMemCmpPass::runOnBlock(BasicBlock &BB, const TargetLibraryInfo *TLI,
852                                   const TargetTransformInfo *TTI,
853                                   const TargetLowering *TL,
854                                   const DataLayout &DL, ProfileSummaryInfo *PSI,
855                                   BlockFrequencyInfo *BFI,
856                                   DomTreeUpdater *DTU) {
857   for (Instruction& I : BB) {
858     CallInst *CI = dyn_cast<CallInst>(&I);
859     if (!CI) {
860       continue;
861     }
862     LibFunc Func;
863     if (TLI->getLibFunc(*CI, Func) &&
864         (Func == LibFunc_memcmp || Func == LibFunc_bcmp) &&
865         expandMemCmp(CI, TTI, TL, &DL, PSI, BFI, DTU)) {
866       return true;
867     }
868   }
869   return false;
870 }
871 
872 PreservedAnalyses
runImpl(Function & F,const TargetLibraryInfo * TLI,const TargetTransformInfo * TTI,const TargetLowering * TL,ProfileSummaryInfo * PSI,BlockFrequencyInfo * BFI,DominatorTree * DT)873 ExpandMemCmpPass::runImpl(Function &F, const TargetLibraryInfo *TLI,
874                           const TargetTransformInfo *TTI,
875                           const TargetLowering *TL, ProfileSummaryInfo *PSI,
876                           BlockFrequencyInfo *BFI, DominatorTree *DT) {
877   Optional<DomTreeUpdater> DTU;
878   if (DT)
879     DTU.emplace(DT, DomTreeUpdater::UpdateStrategy::Lazy);
880 
881   const DataLayout& DL = F.getParent()->getDataLayout();
882   bool MadeChanges = false;
883   for (auto BBIt = F.begin(); BBIt != F.end();) {
884     if (runOnBlock(*BBIt, TLI, TTI, TL, DL, PSI, BFI,
885                    DTU.hasValue() ? DTU.getPointer() : nullptr)) {
886       MadeChanges = true;
887       // If changes were made, restart the function from the beginning, since
888       // the structure of the function was changed.
889       BBIt = F.begin();
890     } else {
891       ++BBIt;
892     }
893   }
894   if (MadeChanges)
895     for (BasicBlock &BB : F)
896       SimplifyInstructionsInBlock(&BB);
897   if (!MadeChanges)
898     return PreservedAnalyses::all();
899   PreservedAnalyses PA;
900   PA.preserve<DominatorTreeAnalysis>();
901   return PA;
902 }
903 
904 } // namespace
905 
906 char ExpandMemCmpPass::ID = 0;
907 INITIALIZE_PASS_BEGIN(ExpandMemCmpPass, "expandmemcmp",
908                       "Expand memcmp() to load/stores", false, false)
INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)909 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
910 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
911 INITIALIZE_PASS_DEPENDENCY(LazyBlockFrequencyInfoPass)
912 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
913 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
914 INITIALIZE_PASS_END(ExpandMemCmpPass, "expandmemcmp",
915                     "Expand memcmp() to load/stores", false, false)
916 
917 FunctionPass *llvm::createExpandMemCmpPass() {
918   return new ExpandMemCmpPass();
919 }
920