xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/MergeICmps.cpp (revision d30a1689f5b37e78ea189232a8b94a7011dc0dc8)
1 //===- MergeICmps.cpp - Optimize chains of integer comparisons ------------===//
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 turns chains of integer comparisons into memcmp (the memcmp is
10 // later typically inlined as a chain of efficient hardware comparisons). This
11 // typically benefits c++ member or nonmember operator==().
12 //
13 // The basic idea is to replace a longer chain of integer comparisons loaded
14 // from contiguous memory locations into a shorter chain of larger integer
15 // comparisons. Benefits are double:
16 //  - There are less jumps, and therefore less opportunities for mispredictions
17 //    and I-cache misses.
18 //  - Code size is smaller, both because jumps are removed and because the
19 //    encoding of a 2*n byte compare is smaller than that of two n-byte
20 //    compares.
21 //
22 // Example:
23 //
24 //  struct S {
25 //    int a;
26 //    char b;
27 //    char c;
28 //    uint16_t d;
29 //    bool operator==(const S& o) const {
30 //      return a == o.a && b == o.b && c == o.c && d == o.d;
31 //    }
32 //  };
33 //
34 //  Is optimized as :
35 //
36 //    bool S::operator==(const S& o) const {
37 //      return memcmp(this, &o, 8) == 0;
38 //    }
39 //
40 //  Which will later be expanded (ExpandMemCmp) as a single 8-bytes icmp.
41 //
42 //===----------------------------------------------------------------------===//
43 
44 #include "llvm/Transforms/Scalar/MergeICmps.h"
45 #include "llvm/Analysis/DomTreeUpdater.h"
46 #include "llvm/Analysis/GlobalsModRef.h"
47 #include "llvm/Analysis/Loads.h"
48 #include "llvm/Analysis/TargetLibraryInfo.h"
49 #include "llvm/Analysis/TargetTransformInfo.h"
50 #include "llvm/IR/Dominators.h"
51 #include "llvm/IR/Function.h"
52 #include "llvm/IR/IRBuilder.h"
53 #include "llvm/InitializePasses.h"
54 #include "llvm/Pass.h"
55 #include "llvm/Transforms/Scalar.h"
56 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
57 #include "llvm/Transforms/Utils/BuildLibCalls.h"
58 #include <algorithm>
59 #include <numeric>
60 #include <utility>
61 #include <vector>
62 
63 using namespace llvm;
64 
65 namespace {
66 
67 #define DEBUG_TYPE "mergeicmps"
68 
69 // A BCE atom "Binary Compare Expression Atom" represents an integer load
70 // that is a constant offset from a base value, e.g. `a` or `o.c` in the example
71 // at the top.
72 struct BCEAtom {
73   BCEAtom() = default;
74   BCEAtom(GetElementPtrInst *GEP, LoadInst *LoadI, int BaseId, APInt Offset)
75       : GEP(GEP), LoadI(LoadI), BaseId(BaseId), Offset(Offset) {}
76 
77   BCEAtom(const BCEAtom &) = delete;
78   BCEAtom &operator=(const BCEAtom &) = delete;
79 
80   BCEAtom(BCEAtom &&that) = default;
81   BCEAtom &operator=(BCEAtom &&that) {
82     if (this == &that)
83       return *this;
84     GEP = that.GEP;
85     LoadI = that.LoadI;
86     BaseId = that.BaseId;
87     Offset = std::move(that.Offset);
88     return *this;
89   }
90 
91   // We want to order BCEAtoms by (Base, Offset). However we cannot use
92   // the pointer values for Base because these are non-deterministic.
93   // To make sure that the sort order is stable, we first assign to each atom
94   // base value an index based on its order of appearance in the chain of
95   // comparisons. We call this index `BaseOrdering`. For example, for:
96   //    b[3] == c[2] && a[1] == d[1] && b[4] == c[3]
97   //    |  block 1 |    |  block 2 |    |  block 3 |
98   // b gets assigned index 0 and a index 1, because b appears as LHS in block 1,
99   // which is before block 2.
100   // We then sort by (BaseOrdering[LHS.Base()], LHS.Offset), which is stable.
101   bool operator<(const BCEAtom &O) const {
102     return BaseId != O.BaseId ? BaseId < O.BaseId : Offset.slt(O.Offset);
103   }
104 
105   GetElementPtrInst *GEP = nullptr;
106   LoadInst *LoadI = nullptr;
107   unsigned BaseId = 0;
108   APInt Offset;
109 };
110 
111 // A class that assigns increasing ids to values in the order in which they are
112 // seen. See comment in `BCEAtom::operator<()``.
113 class BaseIdentifier {
114 public:
115   // Returns the id for value `Base`, after assigning one if `Base` has not been
116   // seen before.
117   int getBaseId(const Value *Base) {
118     assert(Base && "invalid base");
119     const auto Insertion = BaseToIndex.try_emplace(Base, Order);
120     if (Insertion.second)
121       ++Order;
122     return Insertion.first->second;
123   }
124 
125 private:
126   unsigned Order = 1;
127   DenseMap<const Value*, int> BaseToIndex;
128 };
129 
130 // If this value is a load from a constant offset w.r.t. a base address, and
131 // there are no other users of the load or address, returns the base address and
132 // the offset.
133 BCEAtom visitICmpLoadOperand(Value *const Val, BaseIdentifier &BaseId) {
134   auto *const LoadI = dyn_cast<LoadInst>(Val);
135   if (!LoadI)
136     return {};
137   LLVM_DEBUG(dbgs() << "load\n");
138   if (LoadI->isUsedOutsideOfBlock(LoadI->getParent())) {
139     LLVM_DEBUG(dbgs() << "used outside of block\n");
140     return {};
141   }
142   // Do not optimize atomic loads to non-atomic memcmp
143   if (!LoadI->isSimple()) {
144     LLVM_DEBUG(dbgs() << "volatile or atomic\n");
145     return {};
146   }
147   Value *const Addr = LoadI->getOperand(0);
148   if (Addr->getType()->getPointerAddressSpace() != 0) {
149     LLVM_DEBUG(dbgs() << "from non-zero AddressSpace\n");
150     return {};
151   }
152   auto *const GEP = dyn_cast<GetElementPtrInst>(Addr);
153   if (!GEP)
154     return {};
155   LLVM_DEBUG(dbgs() << "GEP\n");
156   if (GEP->isUsedOutsideOfBlock(LoadI->getParent())) {
157     LLVM_DEBUG(dbgs() << "used outside of block\n");
158     return {};
159   }
160   const auto &DL = GEP->getModule()->getDataLayout();
161   if (!isDereferenceablePointer(GEP, LoadI->getType(), DL)) {
162     LLVM_DEBUG(dbgs() << "not dereferenceable\n");
163     // We need to make sure that we can do comparison in any order, so we
164     // require memory to be unconditionnally dereferencable.
165     return {};
166   }
167   APInt Offset = APInt(DL.getPointerTypeSizeInBits(GEP->getType()), 0);
168   if (!GEP->accumulateConstantOffset(DL, Offset))
169     return {};
170   return BCEAtom(GEP, LoadI, BaseId.getBaseId(GEP->getPointerOperand()),
171                  Offset);
172 }
173 
174 // A comparison between two BCE atoms, e.g. `a == o.a` in the example at the
175 // top.
176 // Note: the terminology is misleading: the comparison is symmetric, so there
177 // is no real {l/r}hs. What we want though is to have the same base on the
178 // left (resp. right), so that we can detect consecutive loads. To ensure this
179 // we put the smallest atom on the left.
180 struct BCECmp {
181   BCEAtom Lhs;
182   BCEAtom Rhs;
183   int SizeBits;
184   const ICmpInst *CmpI;
185 
186   BCECmp(BCEAtom L, BCEAtom R, int SizeBits, const ICmpInst *CmpI)
187       : Lhs(std::move(L)), Rhs(std::move(R)), SizeBits(SizeBits), CmpI(CmpI) {
188     if (Rhs < Lhs) std::swap(Rhs, Lhs);
189   }
190 };
191 
192 // A basic block with a comparison between two BCE atoms.
193 // The block might do extra work besides the atom comparison, in which case
194 // doesOtherWork() returns true. Under some conditions, the block can be
195 // split into the atom comparison part and the "other work" part
196 // (see canSplit()).
197 class BCECmpBlock {
198  public:
199   typedef SmallDenseSet<const Instruction *, 8> InstructionSet;
200 
201   BCECmpBlock(BCECmp Cmp, BasicBlock *BB, InstructionSet BlockInsts)
202       : BB(BB), BlockInsts(std::move(BlockInsts)), Cmp(std::move(Cmp)) {}
203 
204   const BCEAtom &Lhs() const { return Cmp.Lhs; }
205   const BCEAtom &Rhs() const { return Cmp.Rhs; }
206   int SizeBits() const { return Cmp.SizeBits; }
207 
208   // Returns true if the block does other works besides comparison.
209   bool doesOtherWork() const;
210 
211   // Returns true if the non-BCE-cmp instructions can be separated from BCE-cmp
212   // instructions in the block.
213   bool canSplit(AliasAnalysis &AA) const;
214 
215   // Return true if this all the relevant instructions in the BCE-cmp-block can
216   // be sunk below this instruction. By doing this, we know we can separate the
217   // BCE-cmp-block instructions from the non-BCE-cmp-block instructions in the
218   // block.
219   bool canSinkBCECmpInst(const Instruction *, AliasAnalysis &AA) const;
220 
221   // We can separate the BCE-cmp-block instructions and the non-BCE-cmp-block
222   // instructions. Split the old block and move all non-BCE-cmp-insts into the
223   // new parent block.
224   void split(BasicBlock *NewParent, AliasAnalysis &AA) const;
225 
226   // The basic block where this comparison happens.
227   BasicBlock *BB;
228   // Instructions relating to the BCECmp and branch.
229   InstructionSet BlockInsts;
230   // The block requires splitting.
231   bool RequireSplit = false;
232   // Original order of this block in the chain.
233   unsigned OrigOrder = 0;
234 
235 private:
236   BCECmp Cmp;
237 };
238 
239 bool BCECmpBlock::canSinkBCECmpInst(const Instruction *Inst,
240                                     AliasAnalysis &AA) const {
241   // If this instruction may clobber the loads and is in middle of the BCE cmp
242   // block instructions, then bail for now.
243   if (Inst->mayWriteToMemory()) {
244     auto MayClobber = [&](LoadInst *LI) {
245       // If a potentially clobbering instruction comes before the load,
246       // we can still safely sink the load.
247       return !Inst->comesBefore(LI) &&
248              isModSet(AA.getModRefInfo(Inst, MemoryLocation::get(LI)));
249     };
250     if (MayClobber(Cmp.Lhs.LoadI) || MayClobber(Cmp.Rhs.LoadI))
251       return false;
252   }
253   // Make sure this instruction does not use any of the BCE cmp block
254   // instructions as operand.
255   return llvm::none_of(Inst->operands(), [&](const Value *Op) {
256     const Instruction *OpI = dyn_cast<Instruction>(Op);
257     return OpI && BlockInsts.contains(OpI);
258   });
259 }
260 
261 void BCECmpBlock::split(BasicBlock *NewParent, AliasAnalysis &AA) const {
262   llvm::SmallVector<Instruction *, 4> OtherInsts;
263   for (Instruction &Inst : *BB) {
264     if (BlockInsts.count(&Inst))
265       continue;
266     assert(canSinkBCECmpInst(&Inst, AA) && "Split unsplittable block");
267     // This is a non-BCE-cmp-block instruction. And it can be separated
268     // from the BCE-cmp-block instruction.
269     OtherInsts.push_back(&Inst);
270   }
271 
272   // Do the actual spliting.
273   for (Instruction *Inst : reverse(OtherInsts)) {
274     Inst->moveBefore(&*NewParent->begin());
275   }
276 }
277 
278 bool BCECmpBlock::canSplit(AliasAnalysis &AA) const {
279   for (Instruction &Inst : *BB) {
280     if (!BlockInsts.count(&Inst)) {
281       if (!canSinkBCECmpInst(&Inst, AA))
282         return false;
283     }
284   }
285   return true;
286 }
287 
288 bool BCECmpBlock::doesOtherWork() const {
289   // TODO(courbet): Can we allow some other things ? This is very conservative.
290   // We might be able to get away with anything does not have any side
291   // effects outside of the basic block.
292   // Note: The GEPs and/or loads are not necessarily in the same block.
293   for (const Instruction &Inst : *BB) {
294     if (!BlockInsts.count(&Inst))
295       return true;
296   }
297   return false;
298 }
299 
300 // Visit the given comparison. If this is a comparison between two valid
301 // BCE atoms, returns the comparison.
302 Optional<BCECmp> visitICmp(const ICmpInst *const CmpI,
303                            const ICmpInst::Predicate ExpectedPredicate,
304                            BaseIdentifier &BaseId) {
305   // The comparison can only be used once:
306   //  - For intermediate blocks, as a branch condition.
307   //  - For the final block, as an incoming value for the Phi.
308   // If there are any other uses of the comparison, we cannot merge it with
309   // other comparisons as we would create an orphan use of the value.
310   if (!CmpI->hasOneUse()) {
311     LLVM_DEBUG(dbgs() << "cmp has several uses\n");
312     return None;
313   }
314   if (CmpI->getPredicate() != ExpectedPredicate)
315     return None;
316   LLVM_DEBUG(dbgs() << "cmp "
317                     << (ExpectedPredicate == ICmpInst::ICMP_EQ ? "eq" : "ne")
318                     << "\n");
319   auto Lhs = visitICmpLoadOperand(CmpI->getOperand(0), BaseId);
320   if (!Lhs.BaseId)
321     return None;
322   auto Rhs = visitICmpLoadOperand(CmpI->getOperand(1), BaseId);
323   if (!Rhs.BaseId)
324     return None;
325   const auto &DL = CmpI->getModule()->getDataLayout();
326   return BCECmp(std::move(Lhs), std::move(Rhs),
327                 DL.getTypeSizeInBits(CmpI->getOperand(0)->getType()), CmpI);
328 }
329 
330 // Visit the given comparison block. If this is a comparison between two valid
331 // BCE atoms, returns the comparison.
332 Optional<BCECmpBlock> visitCmpBlock(Value *const Val, BasicBlock *const Block,
333                                     const BasicBlock *const PhiBlock,
334                                     BaseIdentifier &BaseId) {
335   if (Block->empty()) return None;
336   auto *const BranchI = dyn_cast<BranchInst>(Block->getTerminator());
337   if (!BranchI) return None;
338   LLVM_DEBUG(dbgs() << "branch\n");
339   Value *Cond;
340   ICmpInst::Predicate ExpectedPredicate;
341   if (BranchI->isUnconditional()) {
342     // In this case, we expect an incoming value which is the result of the
343     // comparison. This is the last link in the chain of comparisons (note
344     // that this does not mean that this is the last incoming value, blocks
345     // can be reordered).
346     Cond = Val;
347     ExpectedPredicate = ICmpInst::ICMP_EQ;
348   } else {
349     // In this case, we expect a constant incoming value (the comparison is
350     // chained).
351     const auto *const Const = cast<ConstantInt>(Val);
352     LLVM_DEBUG(dbgs() << "const\n");
353     if (!Const->isZero()) return None;
354     LLVM_DEBUG(dbgs() << "false\n");
355     assert(BranchI->getNumSuccessors() == 2 && "expecting a cond branch");
356     BasicBlock *const FalseBlock = BranchI->getSuccessor(1);
357     Cond = BranchI->getCondition();
358     ExpectedPredicate =
359         FalseBlock == PhiBlock ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
360   }
361 
362   auto *CmpI = dyn_cast<ICmpInst>(Cond);
363   if (!CmpI) return None;
364   LLVM_DEBUG(dbgs() << "icmp\n");
365 
366   Optional<BCECmp> Result = visitICmp(CmpI, ExpectedPredicate, BaseId);
367   if (!Result)
368     return None;
369 
370   BCECmpBlock::InstructionSet BlockInsts(
371       {Result->Lhs.GEP, Result->Rhs.GEP, Result->Lhs.LoadI, Result->Rhs.LoadI,
372        Result->CmpI, BranchI});
373   return BCECmpBlock(std::move(*Result), Block, BlockInsts);
374 }
375 
376 static inline void enqueueBlock(std::vector<BCECmpBlock> &Comparisons,
377                                 BCECmpBlock &&Comparison) {
378   LLVM_DEBUG(dbgs() << "Block '" << Comparison.BB->getName()
379                     << "': Found cmp of " << Comparison.SizeBits()
380                     << " bits between " << Comparison.Lhs().BaseId << " + "
381                     << Comparison.Lhs().Offset << " and "
382                     << Comparison.Rhs().BaseId << " + "
383                     << Comparison.Rhs().Offset << "\n");
384   LLVM_DEBUG(dbgs() << "\n");
385   Comparison.OrigOrder = Comparisons.size();
386   Comparisons.push_back(std::move(Comparison));
387 }
388 
389 // A chain of comparisons.
390 class BCECmpChain {
391 public:
392   using ContiguousBlocks = std::vector<BCECmpBlock>;
393 
394   BCECmpChain(const std::vector<BasicBlock *> &Blocks, PHINode &Phi,
395               AliasAnalysis &AA);
396 
397   bool simplify(const TargetLibraryInfo &TLI, AliasAnalysis &AA,
398                 DomTreeUpdater &DTU);
399 
400   bool atLeastOneMerged() const {
401     return any_of(MergedBlocks_,
402                   [](const auto &Blocks) { return Blocks.size() > 1; });
403   }
404 
405 private:
406   PHINode &Phi_;
407   // The list of all blocks in the chain, grouped by contiguity.
408   std::vector<ContiguousBlocks> MergedBlocks_;
409   // The original entry block (before sorting);
410   BasicBlock *EntryBlock_;
411 };
412 
413 static bool areContiguous(const BCECmpBlock &First, const BCECmpBlock &Second) {
414   return First.Lhs().BaseId == Second.Lhs().BaseId &&
415          First.Rhs().BaseId == Second.Rhs().BaseId &&
416          First.Lhs().Offset + First.SizeBits() / 8 == Second.Lhs().Offset &&
417          First.Rhs().Offset + First.SizeBits() / 8 == Second.Rhs().Offset;
418 }
419 
420 static unsigned getMinOrigOrder(const BCECmpChain::ContiguousBlocks &Blocks) {
421   unsigned MinOrigOrder = std::numeric_limits<unsigned>::max();
422   for (const BCECmpBlock &Block : Blocks)
423     MinOrigOrder = std::min(MinOrigOrder, Block.OrigOrder);
424   return MinOrigOrder;
425 }
426 
427 /// Given a chain of comparison blocks, groups the blocks into contiguous
428 /// ranges that can be merged together into a single comparison.
429 static std::vector<BCECmpChain::ContiguousBlocks>
430 mergeBlocks(std::vector<BCECmpBlock> &&Blocks) {
431   std::vector<BCECmpChain::ContiguousBlocks> MergedBlocks;
432 
433   // Sort to detect continuous offsets.
434   llvm::sort(Blocks,
435              [](const BCECmpBlock &LhsBlock, const BCECmpBlock &RhsBlock) {
436                return std::tie(LhsBlock.Lhs(), LhsBlock.Rhs()) <
437                       std::tie(RhsBlock.Lhs(), RhsBlock.Rhs());
438              });
439 
440   BCECmpChain::ContiguousBlocks *LastMergedBlock = nullptr;
441   for (BCECmpBlock &Block : Blocks) {
442     if (!LastMergedBlock || !areContiguous(LastMergedBlock->back(), Block)) {
443       MergedBlocks.emplace_back();
444       LastMergedBlock = &MergedBlocks.back();
445     } else {
446       LLVM_DEBUG(dbgs() << "Merging block " << Block.BB->getName() << " into "
447                         << LastMergedBlock->back().BB->getName() << "\n");
448     }
449     LastMergedBlock->push_back(std::move(Block));
450   }
451 
452   // While we allow reordering for merging, do not reorder unmerged comparisons.
453   // Doing so may introduce branch on poison.
454   llvm::sort(MergedBlocks, [](const BCECmpChain::ContiguousBlocks &LhsBlocks,
455                               const BCECmpChain::ContiguousBlocks &RhsBlocks) {
456     return getMinOrigOrder(LhsBlocks) < getMinOrigOrder(RhsBlocks);
457   });
458 
459   return MergedBlocks;
460 }
461 
462 BCECmpChain::BCECmpChain(const std::vector<BasicBlock *> &Blocks, PHINode &Phi,
463                          AliasAnalysis &AA)
464     : Phi_(Phi) {
465   assert(!Blocks.empty() && "a chain should have at least one block");
466   // Now look inside blocks to check for BCE comparisons.
467   std::vector<BCECmpBlock> Comparisons;
468   BaseIdentifier BaseId;
469   for (BasicBlock *const Block : Blocks) {
470     assert(Block && "invalid block");
471     Optional<BCECmpBlock> Comparison = visitCmpBlock(
472         Phi.getIncomingValueForBlock(Block), Block, Phi.getParent(), BaseId);
473     if (!Comparison) {
474       LLVM_DEBUG(dbgs() << "chain with invalid BCECmpBlock, no merge.\n");
475       return;
476     }
477     if (Comparison->doesOtherWork()) {
478       LLVM_DEBUG(dbgs() << "block '" << Comparison->BB->getName()
479                         << "' does extra work besides compare\n");
480       if (Comparisons.empty()) {
481         // This is the initial block in the chain, in case this block does other
482         // work, we can try to split the block and move the irrelevant
483         // instructions to the predecessor.
484         //
485         // If this is not the initial block in the chain, splitting it wont
486         // work.
487         //
488         // As once split, there will still be instructions before the BCE cmp
489         // instructions that do other work in program order, i.e. within the
490         // chain before sorting. Unless we can abort the chain at this point
491         // and start anew.
492         //
493         // NOTE: we only handle blocks a with single predecessor for now.
494         if (Comparison->canSplit(AA)) {
495           LLVM_DEBUG(dbgs()
496                      << "Split initial block '" << Comparison->BB->getName()
497                      << "' that does extra work besides compare\n");
498           Comparison->RequireSplit = true;
499           enqueueBlock(Comparisons, std::move(*Comparison));
500         } else {
501           LLVM_DEBUG(dbgs()
502                      << "ignoring initial block '" << Comparison->BB->getName()
503                      << "' that does extra work besides compare\n");
504         }
505         continue;
506       }
507       // TODO(courbet): Right now we abort the whole chain. We could be
508       // merging only the blocks that don't do other work and resume the
509       // chain from there. For example:
510       //  if (a[0] == b[0]) {  // bb1
511       //    if (a[1] == b[1]) {  // bb2
512       //      some_value = 3; //bb3
513       //      if (a[2] == b[2]) { //bb3
514       //        do a ton of stuff  //bb4
515       //      }
516       //    }
517       //  }
518       //
519       // This is:
520       //
521       // bb1 --eq--> bb2 --eq--> bb3* -eq--> bb4 --+
522       //  \            \           \               \
523       //   ne           ne          ne              \
524       //    \            \           \               v
525       //     +------------+-----------+----------> bb_phi
526       //
527       // We can only merge the first two comparisons, because bb3* does
528       // "other work" (setting some_value to 3).
529       // We could still merge bb1 and bb2 though.
530       return;
531     }
532     enqueueBlock(Comparisons, std::move(*Comparison));
533   }
534 
535   // It is possible we have no suitable comparison to merge.
536   if (Comparisons.empty()) {
537     LLVM_DEBUG(dbgs() << "chain with no BCE basic blocks, no merge\n");
538     return;
539   }
540   EntryBlock_ = Comparisons[0].BB;
541   MergedBlocks_ = mergeBlocks(std::move(Comparisons));
542 }
543 
544 namespace {
545 
546 // A class to compute the name of a set of merged basic blocks.
547 // This is optimized for the common case of no block names.
548 class MergedBlockName {
549   // Storage for the uncommon case of several named blocks.
550   SmallString<16> Scratch;
551 
552 public:
553   explicit MergedBlockName(ArrayRef<BCECmpBlock> Comparisons)
554       : Name(makeName(Comparisons)) {}
555   const StringRef Name;
556 
557 private:
558   StringRef makeName(ArrayRef<BCECmpBlock> Comparisons) {
559     assert(!Comparisons.empty() && "no basic block");
560     // Fast path: only one block, or no names at all.
561     if (Comparisons.size() == 1)
562       return Comparisons[0].BB->getName();
563     const int size = std::accumulate(Comparisons.begin(), Comparisons.end(), 0,
564                                      [](int i, const BCECmpBlock &Cmp) {
565                                        return i + Cmp.BB->getName().size();
566                                      });
567     if (size == 0)
568       return StringRef("", 0);
569 
570     // Slow path: at least two blocks, at least one block with a name.
571     Scratch.clear();
572     // We'll have `size` bytes for name and `Comparisons.size() - 1` bytes for
573     // separators.
574     Scratch.reserve(size + Comparisons.size() - 1);
575     const auto append = [this](StringRef str) {
576       Scratch.append(str.begin(), str.end());
577     };
578     append(Comparisons[0].BB->getName());
579     for (int I = 1, E = Comparisons.size(); I < E; ++I) {
580       const BasicBlock *const BB = Comparisons[I].BB;
581       if (!BB->getName().empty()) {
582         append("+");
583         append(BB->getName());
584       }
585     }
586     return Scratch.str();
587   }
588 };
589 } // namespace
590 
591 // Merges the given contiguous comparison blocks into one memcmp block.
592 static BasicBlock *mergeComparisons(ArrayRef<BCECmpBlock> Comparisons,
593                                     BasicBlock *const InsertBefore,
594                                     BasicBlock *const NextCmpBlock,
595                                     PHINode &Phi, const TargetLibraryInfo &TLI,
596                                     AliasAnalysis &AA, DomTreeUpdater &DTU) {
597   assert(!Comparisons.empty() && "merging zero comparisons");
598   LLVMContext &Context = NextCmpBlock->getContext();
599   const BCECmpBlock &FirstCmp = Comparisons[0];
600 
601   // Create a new cmp block before next cmp block.
602   BasicBlock *const BB =
603       BasicBlock::Create(Context, MergedBlockName(Comparisons).Name,
604                          NextCmpBlock->getParent(), InsertBefore);
605   IRBuilder<> Builder(BB);
606   // Add the GEPs from the first BCECmpBlock.
607   Value *const Lhs = Builder.Insert(FirstCmp.Lhs().GEP->clone());
608   Value *const Rhs = Builder.Insert(FirstCmp.Rhs().GEP->clone());
609 
610   Value *IsEqual = nullptr;
611   LLVM_DEBUG(dbgs() << "Merging " << Comparisons.size() << " comparisons -> "
612                     << BB->getName() << "\n");
613 
614   // If there is one block that requires splitting, we do it now, i.e.
615   // just before we know we will collapse the chain. The instructions
616   // can be executed before any of the instructions in the chain.
617   const auto ToSplit = llvm::find_if(
618       Comparisons, [](const BCECmpBlock &B) { return B.RequireSplit; });
619   if (ToSplit != Comparisons.end()) {
620     LLVM_DEBUG(dbgs() << "Splitting non_BCE work to header\n");
621     ToSplit->split(BB, AA);
622   }
623 
624   if (Comparisons.size() == 1) {
625     LLVM_DEBUG(dbgs() << "Only one comparison, updating branches\n");
626     Value *const LhsLoad =
627         Builder.CreateLoad(FirstCmp.Lhs().LoadI->getType(), Lhs);
628     Value *const RhsLoad =
629         Builder.CreateLoad(FirstCmp.Rhs().LoadI->getType(), Rhs);
630     // There are no blocks to merge, just do the comparison.
631     IsEqual = Builder.CreateICmpEQ(LhsLoad, RhsLoad);
632   } else {
633     const unsigned TotalSizeBits = std::accumulate(
634         Comparisons.begin(), Comparisons.end(), 0u,
635         [](int Size, const BCECmpBlock &C) { return Size + C.SizeBits(); });
636 
637     // Create memcmp() == 0.
638     const auto &DL = Phi.getModule()->getDataLayout();
639     Value *const MemCmpCall = emitMemCmp(
640         Lhs, Rhs,
641         ConstantInt::get(DL.getIntPtrType(Context), TotalSizeBits / 8), Builder,
642         DL, &TLI);
643     IsEqual = Builder.CreateICmpEQ(
644         MemCmpCall, ConstantInt::get(Type::getInt32Ty(Context), 0));
645   }
646 
647   BasicBlock *const PhiBB = Phi.getParent();
648   // Add a branch to the next basic block in the chain.
649   if (NextCmpBlock == PhiBB) {
650     // Continue to phi, passing it the comparison result.
651     Builder.CreateBr(PhiBB);
652     Phi.addIncoming(IsEqual, BB);
653     DTU.applyUpdates({{DominatorTree::Insert, BB, PhiBB}});
654   } else {
655     // Continue to next block if equal, exit to phi else.
656     Builder.CreateCondBr(IsEqual, NextCmpBlock, PhiBB);
657     Phi.addIncoming(ConstantInt::getFalse(Context), BB);
658     DTU.applyUpdates({{DominatorTree::Insert, BB, NextCmpBlock},
659                       {DominatorTree::Insert, BB, PhiBB}});
660   }
661   return BB;
662 }
663 
664 bool BCECmpChain::simplify(const TargetLibraryInfo &TLI, AliasAnalysis &AA,
665                            DomTreeUpdater &DTU) {
666   assert(atLeastOneMerged() && "simplifying trivial BCECmpChain");
667   LLVM_DEBUG(dbgs() << "Simplifying comparison chain starting at block "
668                     << EntryBlock_->getName() << "\n");
669 
670   // Effectively merge blocks. We go in the reverse direction from the phi block
671   // so that the next block is always available to branch to.
672   BasicBlock *InsertBefore = EntryBlock_;
673   BasicBlock *NextCmpBlock = Phi_.getParent();
674   for (const auto &Blocks : reverse(MergedBlocks_)) {
675     InsertBefore = NextCmpBlock = mergeComparisons(
676         Blocks, InsertBefore, NextCmpBlock, Phi_, TLI, AA, DTU);
677   }
678 
679   // Replace the original cmp chain with the new cmp chain by pointing all
680   // predecessors of EntryBlock_ to NextCmpBlock instead. This makes all cmp
681   // blocks in the old chain unreachable.
682   while (!pred_empty(EntryBlock_)) {
683     BasicBlock* const Pred = *pred_begin(EntryBlock_);
684     LLVM_DEBUG(dbgs() << "Updating jump into old chain from " << Pred->getName()
685                       << "\n");
686     Pred->getTerminator()->replaceUsesOfWith(EntryBlock_, NextCmpBlock);
687     DTU.applyUpdates({{DominatorTree::Delete, Pred, EntryBlock_},
688                       {DominatorTree::Insert, Pred, NextCmpBlock}});
689   }
690 
691   // If the old cmp chain was the function entry, we need to update the function
692   // entry.
693   const bool ChainEntryIsFnEntry = EntryBlock_->isEntryBlock();
694   if (ChainEntryIsFnEntry && DTU.hasDomTree()) {
695     LLVM_DEBUG(dbgs() << "Changing function entry from "
696                       << EntryBlock_->getName() << " to "
697                       << NextCmpBlock->getName() << "\n");
698     DTU.getDomTree().setNewRoot(NextCmpBlock);
699     DTU.applyUpdates({{DominatorTree::Delete, NextCmpBlock, EntryBlock_}});
700   }
701   EntryBlock_ = nullptr;
702 
703   // Delete merged blocks. This also removes incoming values in phi.
704   SmallVector<BasicBlock *, 16> DeadBlocks;
705   for (const auto &Blocks : MergedBlocks_) {
706     for (const BCECmpBlock &Block : Blocks) {
707       LLVM_DEBUG(dbgs() << "Deleting merged block " << Block.BB->getName()
708                         << "\n");
709       DeadBlocks.push_back(Block.BB);
710     }
711   }
712   DeleteDeadBlocks(DeadBlocks, &DTU);
713 
714   MergedBlocks_.clear();
715   return true;
716 }
717 
718 std::vector<BasicBlock *> getOrderedBlocks(PHINode &Phi,
719                                            BasicBlock *const LastBlock,
720                                            int NumBlocks) {
721   // Walk up from the last block to find other blocks.
722   std::vector<BasicBlock *> Blocks(NumBlocks);
723   assert(LastBlock && "invalid last block");
724   BasicBlock *CurBlock = LastBlock;
725   for (int BlockIndex = NumBlocks - 1; BlockIndex > 0; --BlockIndex) {
726     if (CurBlock->hasAddressTaken()) {
727       // Somebody is jumping to the block through an address, all bets are
728       // off.
729       LLVM_DEBUG(dbgs() << "skip: block " << BlockIndex
730                         << " has its address taken\n");
731       return {};
732     }
733     Blocks[BlockIndex] = CurBlock;
734     auto *SinglePredecessor = CurBlock->getSinglePredecessor();
735     if (!SinglePredecessor) {
736       // The block has two or more predecessors.
737       LLVM_DEBUG(dbgs() << "skip: block " << BlockIndex
738                         << " has two or more predecessors\n");
739       return {};
740     }
741     if (Phi.getBasicBlockIndex(SinglePredecessor) < 0) {
742       // The block does not link back to the phi.
743       LLVM_DEBUG(dbgs() << "skip: block " << BlockIndex
744                         << " does not link back to the phi\n");
745       return {};
746     }
747     CurBlock = SinglePredecessor;
748   }
749   Blocks[0] = CurBlock;
750   return Blocks;
751 }
752 
753 bool processPhi(PHINode &Phi, const TargetLibraryInfo &TLI, AliasAnalysis &AA,
754                 DomTreeUpdater &DTU) {
755   LLVM_DEBUG(dbgs() << "processPhi()\n");
756   if (Phi.getNumIncomingValues() <= 1) {
757     LLVM_DEBUG(dbgs() << "skip: only one incoming value in phi\n");
758     return false;
759   }
760   // We are looking for something that has the following structure:
761   //   bb1 --eq--> bb2 --eq--> bb3 --eq--> bb4 --+
762   //     \            \           \               \
763   //      ne           ne          ne              \
764   //       \            \           \               v
765   //        +------------+-----------+----------> bb_phi
766   //
767   //  - The last basic block (bb4 here) must branch unconditionally to bb_phi.
768   //    It's the only block that contributes a non-constant value to the Phi.
769   //  - All other blocks (b1, b2, b3) must have exactly two successors, one of
770   //    them being the phi block.
771   //  - All intermediate blocks (bb2, bb3) must have only one predecessor.
772   //  - Blocks cannot do other work besides the comparison, see doesOtherWork()
773 
774   // The blocks are not necessarily ordered in the phi, so we start from the
775   // last block and reconstruct the order.
776   BasicBlock *LastBlock = nullptr;
777   for (unsigned I = 0; I < Phi.getNumIncomingValues(); ++I) {
778     if (isa<ConstantInt>(Phi.getIncomingValue(I))) continue;
779     if (LastBlock) {
780       // There are several non-constant values.
781       LLVM_DEBUG(dbgs() << "skip: several non-constant values\n");
782       return false;
783     }
784     if (!isa<ICmpInst>(Phi.getIncomingValue(I)) ||
785         cast<ICmpInst>(Phi.getIncomingValue(I))->getParent() !=
786             Phi.getIncomingBlock(I)) {
787       // Non-constant incoming value is not from a cmp instruction or not
788       // produced by the last block. We could end up processing the value
789       // producing block more than once.
790       //
791       // This is an uncommon case, so we bail.
792       LLVM_DEBUG(
793           dbgs()
794           << "skip: non-constant value not from cmp or not from last block.\n");
795       return false;
796     }
797     LastBlock = Phi.getIncomingBlock(I);
798   }
799   if (!LastBlock) {
800     // There is no non-constant block.
801     LLVM_DEBUG(dbgs() << "skip: no non-constant block\n");
802     return false;
803   }
804   if (LastBlock->getSingleSuccessor() != Phi.getParent()) {
805     LLVM_DEBUG(dbgs() << "skip: last block non-phi successor\n");
806     return false;
807   }
808 
809   const auto Blocks =
810       getOrderedBlocks(Phi, LastBlock, Phi.getNumIncomingValues());
811   if (Blocks.empty()) return false;
812   BCECmpChain CmpChain(Blocks, Phi, AA);
813 
814   if (!CmpChain.atLeastOneMerged()) {
815     LLVM_DEBUG(dbgs() << "skip: nothing merged\n");
816     return false;
817   }
818 
819   return CmpChain.simplify(TLI, AA, DTU);
820 }
821 
822 static bool runImpl(Function &F, const TargetLibraryInfo &TLI,
823                     const TargetTransformInfo &TTI, AliasAnalysis &AA,
824                     DominatorTree *DT) {
825   LLVM_DEBUG(dbgs() << "MergeICmpsLegacyPass: " << F.getName() << "\n");
826 
827   // We only try merging comparisons if the target wants to expand memcmp later.
828   // The rationale is to avoid turning small chains into memcmp calls.
829   if (!TTI.enableMemCmpExpansion(F.hasOptSize(), true))
830     return false;
831 
832   // If we don't have memcmp avaiable we can't emit calls to it.
833   if (!TLI.has(LibFunc_memcmp))
834     return false;
835 
836   DomTreeUpdater DTU(DT, /*PostDominatorTree*/ nullptr,
837                      DomTreeUpdater::UpdateStrategy::Eager);
838 
839   bool MadeChange = false;
840 
841   for (BasicBlock &BB : llvm::drop_begin(F)) {
842     // A Phi operation is always first in a basic block.
843     if (auto *const Phi = dyn_cast<PHINode>(&*BB.begin()))
844       MadeChange |= processPhi(*Phi, TLI, AA, DTU);
845   }
846 
847   return MadeChange;
848 }
849 
850 class MergeICmpsLegacyPass : public FunctionPass {
851 public:
852   static char ID;
853 
854   MergeICmpsLegacyPass() : FunctionPass(ID) {
855     initializeMergeICmpsLegacyPassPass(*PassRegistry::getPassRegistry());
856   }
857 
858   bool runOnFunction(Function &F) override {
859     if (skipFunction(F)) return false;
860     const auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
861     const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
862     // MergeICmps does not need the DominatorTree, but we update it if it's
863     // already available.
864     auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
865     auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
866     return runImpl(F, TLI, TTI, AA, DTWP ? &DTWP->getDomTree() : nullptr);
867   }
868 
869  private:
870   void getAnalysisUsage(AnalysisUsage &AU) const override {
871     AU.addRequired<TargetLibraryInfoWrapperPass>();
872     AU.addRequired<TargetTransformInfoWrapperPass>();
873     AU.addRequired<AAResultsWrapperPass>();
874     AU.addPreserved<GlobalsAAWrapperPass>();
875     AU.addPreserved<DominatorTreeWrapperPass>();
876   }
877 };
878 
879 } // namespace
880 
881 char MergeICmpsLegacyPass::ID = 0;
882 INITIALIZE_PASS_BEGIN(MergeICmpsLegacyPass, "mergeicmps",
883                       "Merge contiguous icmps into a memcmp", false, false)
884 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
885 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
886 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
887 INITIALIZE_PASS_END(MergeICmpsLegacyPass, "mergeicmps",
888                     "Merge contiguous icmps into a memcmp", false, false)
889 
890 Pass *llvm::createMergeICmpsLegacyPass() { return new MergeICmpsLegacyPass(); }
891 
892 PreservedAnalyses MergeICmpsPass::run(Function &F,
893                                       FunctionAnalysisManager &AM) {
894   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
895   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
896   auto &AA = AM.getResult<AAManager>(F);
897   auto *DT = AM.getCachedResult<DominatorTreeAnalysis>(F);
898   const bool MadeChanges = runImpl(F, TLI, TTI, AA, DT);
899   if (!MadeChanges)
900     return PreservedAnalyses::all();
901   PreservedAnalyses PA;
902   PA.preserve<DominatorTreeAnalysis>();
903   return PA;
904 }
905