xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/MergeICmps.cpp (revision e6bfd18d21b225af6a0ed67ceeaf1293b7b9eba5)
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 *Addr = LoadI->getOperand(0);
148   if (Addr->getType()->getPointerAddressSpace() != 0) {
149     LLVM_DEBUG(dbgs() << "from non-zero AddressSpace\n");
150     return {};
151   }
152   const auto &DL = LoadI->getModule()->getDataLayout();
153   if (!isDereferenceablePointer(Addr, LoadI->getType(), DL)) {
154     LLVM_DEBUG(dbgs() << "not dereferenceable\n");
155     // We need to make sure that we can do comparison in any order, so we
156     // require memory to be unconditionnally dereferencable.
157     return {};
158   }
159 
160   APInt Offset = APInt(DL.getPointerTypeSizeInBits(Addr->getType()), 0);
161   Value *Base = Addr;
162   auto *GEP = dyn_cast<GetElementPtrInst>(Addr);
163   if (GEP) {
164     LLVM_DEBUG(dbgs() << "GEP\n");
165     if (GEP->isUsedOutsideOfBlock(LoadI->getParent())) {
166       LLVM_DEBUG(dbgs() << "used outside of block\n");
167       return {};
168     }
169     if (!GEP->accumulateConstantOffset(DL, Offset))
170       return {};
171     Base = GEP->getPointerOperand();
172   }
173   return BCEAtom(GEP, LoadI, BaseId.getBaseId(Base), Offset);
174 }
175 
176 // A comparison between two BCE atoms, e.g. `a == o.a` in the example at the
177 // top.
178 // Note: the terminology is misleading: the comparison is symmetric, so there
179 // is no real {l/r}hs. What we want though is to have the same base on the
180 // left (resp. right), so that we can detect consecutive loads. To ensure this
181 // we put the smallest atom on the left.
182 struct BCECmp {
183   BCEAtom Lhs;
184   BCEAtom Rhs;
185   int SizeBits;
186   const ICmpInst *CmpI;
187 
188   BCECmp(BCEAtom L, BCEAtom R, int SizeBits, const ICmpInst *CmpI)
189       : Lhs(std::move(L)), Rhs(std::move(R)), SizeBits(SizeBits), CmpI(CmpI) {
190     if (Rhs < Lhs) std::swap(Rhs, Lhs);
191   }
192 };
193 
194 // A basic block with a comparison between two BCE atoms.
195 // The block might do extra work besides the atom comparison, in which case
196 // doesOtherWork() returns true. Under some conditions, the block can be
197 // split into the atom comparison part and the "other work" part
198 // (see canSplit()).
199 class BCECmpBlock {
200  public:
201   typedef SmallDenseSet<const Instruction *, 8> InstructionSet;
202 
203   BCECmpBlock(BCECmp Cmp, BasicBlock *BB, InstructionSet BlockInsts)
204       : BB(BB), BlockInsts(std::move(BlockInsts)), Cmp(std::move(Cmp)) {}
205 
206   const BCEAtom &Lhs() const { return Cmp.Lhs; }
207   const BCEAtom &Rhs() const { return Cmp.Rhs; }
208   int SizeBits() const { return Cmp.SizeBits; }
209 
210   // Returns true if the block does other works besides comparison.
211   bool doesOtherWork() const;
212 
213   // Returns true if the non-BCE-cmp instructions can be separated from BCE-cmp
214   // instructions in the block.
215   bool canSplit(AliasAnalysis &AA) const;
216 
217   // Return true if this all the relevant instructions in the BCE-cmp-block can
218   // be sunk below this instruction. By doing this, we know we can separate the
219   // BCE-cmp-block instructions from the non-BCE-cmp-block instructions in the
220   // block.
221   bool canSinkBCECmpInst(const Instruction *, AliasAnalysis &AA) const;
222 
223   // We can separate the BCE-cmp-block instructions and the non-BCE-cmp-block
224   // instructions. Split the old block and move all non-BCE-cmp-insts into the
225   // new parent block.
226   void split(BasicBlock *NewParent, AliasAnalysis &AA) const;
227 
228   // The basic block where this comparison happens.
229   BasicBlock *BB;
230   // Instructions relating to the BCECmp and branch.
231   InstructionSet BlockInsts;
232   // The block requires splitting.
233   bool RequireSplit = false;
234   // Original order of this block in the chain.
235   unsigned OrigOrder = 0;
236 
237 private:
238   BCECmp Cmp;
239 };
240 
241 bool BCECmpBlock::canSinkBCECmpInst(const Instruction *Inst,
242                                     AliasAnalysis &AA) const {
243   // If this instruction may clobber the loads and is in middle of the BCE cmp
244   // block instructions, then bail for now.
245   if (Inst->mayWriteToMemory()) {
246     auto MayClobber = [&](LoadInst *LI) {
247       // If a potentially clobbering instruction comes before the load,
248       // we can still safely sink the load.
249       return (Inst->getParent() != LI->getParent() || !Inst->comesBefore(LI)) &&
250              isModSet(AA.getModRefInfo(Inst, MemoryLocation::get(LI)));
251     };
252     if (MayClobber(Cmp.Lhs.LoadI) || MayClobber(Cmp.Rhs.LoadI))
253       return false;
254   }
255   // Make sure this instruction does not use any of the BCE cmp block
256   // instructions as operand.
257   return llvm::none_of(Inst->operands(), [&](const Value *Op) {
258     const Instruction *OpI = dyn_cast<Instruction>(Op);
259     return OpI && BlockInsts.contains(OpI);
260   });
261 }
262 
263 void BCECmpBlock::split(BasicBlock *NewParent, AliasAnalysis &AA) const {
264   llvm::SmallVector<Instruction *, 4> OtherInsts;
265   for (Instruction &Inst : *BB) {
266     if (BlockInsts.count(&Inst))
267       continue;
268     assert(canSinkBCECmpInst(&Inst, AA) && "Split unsplittable block");
269     // This is a non-BCE-cmp-block instruction. And it can be separated
270     // from the BCE-cmp-block instruction.
271     OtherInsts.push_back(&Inst);
272   }
273 
274   // Do the actual spliting.
275   for (Instruction *Inst : reverse(OtherInsts))
276     Inst->moveBefore(*NewParent, NewParent->begin());
277 }
278 
279 bool BCECmpBlock::canSplit(AliasAnalysis &AA) const {
280   for (Instruction &Inst : *BB) {
281     if (!BlockInsts.count(&Inst)) {
282       if (!canSinkBCECmpInst(&Inst, AA))
283         return false;
284     }
285   }
286   return true;
287 }
288 
289 bool BCECmpBlock::doesOtherWork() const {
290   // TODO(courbet): Can we allow some other things ? This is very conservative.
291   // We might be able to get away with anything does not have any side
292   // effects outside of the basic block.
293   // Note: The GEPs and/or loads are not necessarily in the same block.
294   for (const Instruction &Inst : *BB) {
295     if (!BlockInsts.count(&Inst))
296       return true;
297   }
298   return false;
299 }
300 
301 // Visit the given comparison. If this is a comparison between two valid
302 // BCE atoms, returns the comparison.
303 Optional<BCECmp> visitICmp(const ICmpInst *const CmpI,
304                            const ICmpInst::Predicate ExpectedPredicate,
305                            BaseIdentifier &BaseId) {
306   // The comparison can only be used once:
307   //  - For intermediate blocks, as a branch condition.
308   //  - For the final block, as an incoming value for the Phi.
309   // If there are any other uses of the comparison, we cannot merge it with
310   // other comparisons as we would create an orphan use of the value.
311   if (!CmpI->hasOneUse()) {
312     LLVM_DEBUG(dbgs() << "cmp has several uses\n");
313     return None;
314   }
315   if (CmpI->getPredicate() != ExpectedPredicate)
316     return None;
317   LLVM_DEBUG(dbgs() << "cmp "
318                     << (ExpectedPredicate == ICmpInst::ICMP_EQ ? "eq" : "ne")
319                     << "\n");
320   auto Lhs = visitICmpLoadOperand(CmpI->getOperand(0), BaseId);
321   if (!Lhs.BaseId)
322     return None;
323   auto Rhs = visitICmpLoadOperand(CmpI->getOperand(1), BaseId);
324   if (!Rhs.BaseId)
325     return None;
326   const auto &DL = CmpI->getModule()->getDataLayout();
327   return BCECmp(std::move(Lhs), std::move(Rhs),
328                 DL.getTypeSizeInBits(CmpI->getOperand(0)->getType()), CmpI);
329 }
330 
331 // Visit the given comparison block. If this is a comparison between two valid
332 // BCE atoms, returns the comparison.
333 Optional<BCECmpBlock> visitCmpBlock(Value *const Val, BasicBlock *const Block,
334                                     const BasicBlock *const PhiBlock,
335                                     BaseIdentifier &BaseId) {
336   if (Block->empty()) return None;
337   auto *const BranchI = dyn_cast<BranchInst>(Block->getTerminator());
338   if (!BranchI) return None;
339   LLVM_DEBUG(dbgs() << "branch\n");
340   Value *Cond;
341   ICmpInst::Predicate ExpectedPredicate;
342   if (BranchI->isUnconditional()) {
343     // In this case, we expect an incoming value which is the result of the
344     // comparison. This is the last link in the chain of comparisons (note
345     // that this does not mean that this is the last incoming value, blocks
346     // can be reordered).
347     Cond = Val;
348     ExpectedPredicate = ICmpInst::ICMP_EQ;
349   } else {
350     // In this case, we expect a constant incoming value (the comparison is
351     // chained).
352     const auto *const Const = cast<ConstantInt>(Val);
353     LLVM_DEBUG(dbgs() << "const\n");
354     if (!Const->isZero()) return None;
355     LLVM_DEBUG(dbgs() << "false\n");
356     assert(BranchI->getNumSuccessors() == 2 && "expecting a cond branch");
357     BasicBlock *const FalseBlock = BranchI->getSuccessor(1);
358     Cond = BranchI->getCondition();
359     ExpectedPredicate =
360         FalseBlock == PhiBlock ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
361   }
362 
363   auto *CmpI = dyn_cast<ICmpInst>(Cond);
364   if (!CmpI) return None;
365   LLVM_DEBUG(dbgs() << "icmp\n");
366 
367   Optional<BCECmp> Result = visitICmp(CmpI, ExpectedPredicate, BaseId);
368   if (!Result)
369     return None;
370 
371   BCECmpBlock::InstructionSet BlockInsts(
372       {Result->Lhs.LoadI, Result->Rhs.LoadI, Result->CmpI, BranchI});
373   if (Result->Lhs.GEP)
374     BlockInsts.insert(Result->Lhs.GEP);
375   if (Result->Rhs.GEP)
376     BlockInsts.insert(Result->Rhs.GEP);
377   return BCECmpBlock(std::move(*Result), Block, BlockInsts);
378 }
379 
380 static inline void enqueueBlock(std::vector<BCECmpBlock> &Comparisons,
381                                 BCECmpBlock &&Comparison) {
382   LLVM_DEBUG(dbgs() << "Block '" << Comparison.BB->getName()
383                     << "': Found cmp of " << Comparison.SizeBits()
384                     << " bits between " << Comparison.Lhs().BaseId << " + "
385                     << Comparison.Lhs().Offset << " and "
386                     << Comparison.Rhs().BaseId << " + "
387                     << Comparison.Rhs().Offset << "\n");
388   LLVM_DEBUG(dbgs() << "\n");
389   Comparison.OrigOrder = Comparisons.size();
390   Comparisons.push_back(std::move(Comparison));
391 }
392 
393 // A chain of comparisons.
394 class BCECmpChain {
395 public:
396   using ContiguousBlocks = std::vector<BCECmpBlock>;
397 
398   BCECmpChain(const std::vector<BasicBlock *> &Blocks, PHINode &Phi,
399               AliasAnalysis &AA);
400 
401   bool simplify(const TargetLibraryInfo &TLI, AliasAnalysis &AA,
402                 DomTreeUpdater &DTU);
403 
404   bool atLeastOneMerged() const {
405     return any_of(MergedBlocks_,
406                   [](const auto &Blocks) { return Blocks.size() > 1; });
407   }
408 
409 private:
410   PHINode &Phi_;
411   // The list of all blocks in the chain, grouped by contiguity.
412   std::vector<ContiguousBlocks> MergedBlocks_;
413   // The original entry block (before sorting);
414   BasicBlock *EntryBlock_;
415 };
416 
417 static bool areContiguous(const BCECmpBlock &First, const BCECmpBlock &Second) {
418   return First.Lhs().BaseId == Second.Lhs().BaseId &&
419          First.Rhs().BaseId == Second.Rhs().BaseId &&
420          First.Lhs().Offset + First.SizeBits() / 8 == Second.Lhs().Offset &&
421          First.Rhs().Offset + First.SizeBits() / 8 == Second.Rhs().Offset;
422 }
423 
424 static unsigned getMinOrigOrder(const BCECmpChain::ContiguousBlocks &Blocks) {
425   unsigned MinOrigOrder = std::numeric_limits<unsigned>::max();
426   for (const BCECmpBlock &Block : Blocks)
427     MinOrigOrder = std::min(MinOrigOrder, Block.OrigOrder);
428   return MinOrigOrder;
429 }
430 
431 /// Given a chain of comparison blocks, groups the blocks into contiguous
432 /// ranges that can be merged together into a single comparison.
433 static std::vector<BCECmpChain::ContiguousBlocks>
434 mergeBlocks(std::vector<BCECmpBlock> &&Blocks) {
435   std::vector<BCECmpChain::ContiguousBlocks> MergedBlocks;
436 
437   // Sort to detect continuous offsets.
438   llvm::sort(Blocks,
439              [](const BCECmpBlock &LhsBlock, const BCECmpBlock &RhsBlock) {
440                return std::tie(LhsBlock.Lhs(), LhsBlock.Rhs()) <
441                       std::tie(RhsBlock.Lhs(), RhsBlock.Rhs());
442              });
443 
444   BCECmpChain::ContiguousBlocks *LastMergedBlock = nullptr;
445   for (BCECmpBlock &Block : Blocks) {
446     if (!LastMergedBlock || !areContiguous(LastMergedBlock->back(), Block)) {
447       MergedBlocks.emplace_back();
448       LastMergedBlock = &MergedBlocks.back();
449     } else {
450       LLVM_DEBUG(dbgs() << "Merging block " << Block.BB->getName() << " into "
451                         << LastMergedBlock->back().BB->getName() << "\n");
452     }
453     LastMergedBlock->push_back(std::move(Block));
454   }
455 
456   // While we allow reordering for merging, do not reorder unmerged comparisons.
457   // Doing so may introduce branch on poison.
458   llvm::sort(MergedBlocks, [](const BCECmpChain::ContiguousBlocks &LhsBlocks,
459                               const BCECmpChain::ContiguousBlocks &RhsBlocks) {
460     return getMinOrigOrder(LhsBlocks) < getMinOrigOrder(RhsBlocks);
461   });
462 
463   return MergedBlocks;
464 }
465 
466 BCECmpChain::BCECmpChain(const std::vector<BasicBlock *> &Blocks, PHINode &Phi,
467                          AliasAnalysis &AA)
468     : Phi_(Phi) {
469   assert(!Blocks.empty() && "a chain should have at least one block");
470   // Now look inside blocks to check for BCE comparisons.
471   std::vector<BCECmpBlock> Comparisons;
472   BaseIdentifier BaseId;
473   for (BasicBlock *const Block : Blocks) {
474     assert(Block && "invalid block");
475     Optional<BCECmpBlock> Comparison = visitCmpBlock(
476         Phi.getIncomingValueForBlock(Block), Block, Phi.getParent(), BaseId);
477     if (!Comparison) {
478       LLVM_DEBUG(dbgs() << "chain with invalid BCECmpBlock, no merge.\n");
479       return;
480     }
481     if (Comparison->doesOtherWork()) {
482       LLVM_DEBUG(dbgs() << "block '" << Comparison->BB->getName()
483                         << "' does extra work besides compare\n");
484       if (Comparisons.empty()) {
485         // This is the initial block in the chain, in case this block does other
486         // work, we can try to split the block and move the irrelevant
487         // instructions to the predecessor.
488         //
489         // If this is not the initial block in the chain, splitting it wont
490         // work.
491         //
492         // As once split, there will still be instructions before the BCE cmp
493         // instructions that do other work in program order, i.e. within the
494         // chain before sorting. Unless we can abort the chain at this point
495         // and start anew.
496         //
497         // NOTE: we only handle blocks a with single predecessor for now.
498         if (Comparison->canSplit(AA)) {
499           LLVM_DEBUG(dbgs()
500                      << "Split initial block '" << Comparison->BB->getName()
501                      << "' that does extra work besides compare\n");
502           Comparison->RequireSplit = true;
503           enqueueBlock(Comparisons, std::move(*Comparison));
504         } else {
505           LLVM_DEBUG(dbgs()
506                      << "ignoring initial block '" << Comparison->BB->getName()
507                      << "' that does extra work besides compare\n");
508         }
509         continue;
510       }
511       // TODO(courbet): Right now we abort the whole chain. We could be
512       // merging only the blocks that don't do other work and resume the
513       // chain from there. For example:
514       //  if (a[0] == b[0]) {  // bb1
515       //    if (a[1] == b[1]) {  // bb2
516       //      some_value = 3; //bb3
517       //      if (a[2] == b[2]) { //bb3
518       //        do a ton of stuff  //bb4
519       //      }
520       //    }
521       //  }
522       //
523       // This is:
524       //
525       // bb1 --eq--> bb2 --eq--> bb3* -eq--> bb4 --+
526       //  \            \           \               \
527       //   ne           ne          ne              \
528       //    \            \           \               v
529       //     +------------+-----------+----------> bb_phi
530       //
531       // We can only merge the first two comparisons, because bb3* does
532       // "other work" (setting some_value to 3).
533       // We could still merge bb1 and bb2 though.
534       return;
535     }
536     enqueueBlock(Comparisons, std::move(*Comparison));
537   }
538 
539   // It is possible we have no suitable comparison to merge.
540   if (Comparisons.empty()) {
541     LLVM_DEBUG(dbgs() << "chain with no BCE basic blocks, no merge\n");
542     return;
543   }
544   EntryBlock_ = Comparisons[0].BB;
545   MergedBlocks_ = mergeBlocks(std::move(Comparisons));
546 }
547 
548 namespace {
549 
550 // A class to compute the name of a set of merged basic blocks.
551 // This is optimized for the common case of no block names.
552 class MergedBlockName {
553   // Storage for the uncommon case of several named blocks.
554   SmallString<16> Scratch;
555 
556 public:
557   explicit MergedBlockName(ArrayRef<BCECmpBlock> Comparisons)
558       : Name(makeName(Comparisons)) {}
559   const StringRef Name;
560 
561 private:
562   StringRef makeName(ArrayRef<BCECmpBlock> Comparisons) {
563     assert(!Comparisons.empty() && "no basic block");
564     // Fast path: only one block, or no names at all.
565     if (Comparisons.size() == 1)
566       return Comparisons[0].BB->getName();
567     const int size = std::accumulate(Comparisons.begin(), Comparisons.end(), 0,
568                                      [](int i, const BCECmpBlock &Cmp) {
569                                        return i + Cmp.BB->getName().size();
570                                      });
571     if (size == 0)
572       return StringRef("", 0);
573 
574     // Slow path: at least two blocks, at least one block with a name.
575     Scratch.clear();
576     // We'll have `size` bytes for name and `Comparisons.size() - 1` bytes for
577     // separators.
578     Scratch.reserve(size + Comparisons.size() - 1);
579     const auto append = [this](StringRef str) {
580       Scratch.append(str.begin(), str.end());
581     };
582     append(Comparisons[0].BB->getName());
583     for (int I = 1, E = Comparisons.size(); I < E; ++I) {
584       const BasicBlock *const BB = Comparisons[I].BB;
585       if (!BB->getName().empty()) {
586         append("+");
587         append(BB->getName());
588       }
589     }
590     return Scratch.str();
591   }
592 };
593 } // namespace
594 
595 // Merges the given contiguous comparison blocks into one memcmp block.
596 static BasicBlock *mergeComparisons(ArrayRef<BCECmpBlock> Comparisons,
597                                     BasicBlock *const InsertBefore,
598                                     BasicBlock *const NextCmpBlock,
599                                     PHINode &Phi, const TargetLibraryInfo &TLI,
600                                     AliasAnalysis &AA, DomTreeUpdater &DTU) {
601   assert(!Comparisons.empty() && "merging zero comparisons");
602   LLVMContext &Context = NextCmpBlock->getContext();
603   const BCECmpBlock &FirstCmp = Comparisons[0];
604 
605   // Create a new cmp block before next cmp block.
606   BasicBlock *const BB =
607       BasicBlock::Create(Context, MergedBlockName(Comparisons).Name,
608                          NextCmpBlock->getParent(), InsertBefore);
609   IRBuilder<> Builder(BB);
610   // Add the GEPs from the first BCECmpBlock.
611   Value *Lhs, *Rhs;
612   if (FirstCmp.Lhs().GEP)
613     Lhs = Builder.Insert(FirstCmp.Lhs().GEP->clone());
614   else
615     Lhs = FirstCmp.Lhs().LoadI->getPointerOperand();
616   if (FirstCmp.Rhs().GEP)
617     Rhs = Builder.Insert(FirstCmp.Rhs().GEP->clone());
618   else
619     Rhs = FirstCmp.Rhs().LoadI->getPointerOperand();
620 
621   Value *IsEqual = nullptr;
622   LLVM_DEBUG(dbgs() << "Merging " << Comparisons.size() << " comparisons -> "
623                     << BB->getName() << "\n");
624 
625   // If there is one block that requires splitting, we do it now, i.e.
626   // just before we know we will collapse the chain. The instructions
627   // can be executed before any of the instructions in the chain.
628   const auto ToSplit = llvm::find_if(
629       Comparisons, [](const BCECmpBlock &B) { return B.RequireSplit; });
630   if (ToSplit != Comparisons.end()) {
631     LLVM_DEBUG(dbgs() << "Splitting non_BCE work to header\n");
632     ToSplit->split(BB, AA);
633   }
634 
635   if (Comparisons.size() == 1) {
636     LLVM_DEBUG(dbgs() << "Only one comparison, updating branches\n");
637     Value *const LhsLoad =
638         Builder.CreateLoad(FirstCmp.Lhs().LoadI->getType(), Lhs);
639     Value *const RhsLoad =
640         Builder.CreateLoad(FirstCmp.Rhs().LoadI->getType(), Rhs);
641     // There are no blocks to merge, just do the comparison.
642     IsEqual = Builder.CreateICmpEQ(LhsLoad, RhsLoad);
643   } else {
644     const unsigned TotalSizeBits = std::accumulate(
645         Comparisons.begin(), Comparisons.end(), 0u,
646         [](int Size, const BCECmpBlock &C) { return Size + C.SizeBits(); });
647 
648     // Create memcmp() == 0.
649     const auto &DL = Phi.getModule()->getDataLayout();
650     Value *const MemCmpCall = emitMemCmp(
651         Lhs, Rhs,
652         ConstantInt::get(DL.getIntPtrType(Context), TotalSizeBits / 8), Builder,
653         DL, &TLI);
654     IsEqual = Builder.CreateICmpEQ(
655         MemCmpCall, ConstantInt::get(Type::getInt32Ty(Context), 0));
656   }
657 
658   BasicBlock *const PhiBB = Phi.getParent();
659   // Add a branch to the next basic block in the chain.
660   if (NextCmpBlock == PhiBB) {
661     // Continue to phi, passing it the comparison result.
662     Builder.CreateBr(PhiBB);
663     Phi.addIncoming(IsEqual, BB);
664     DTU.applyUpdates({{DominatorTree::Insert, BB, PhiBB}});
665   } else {
666     // Continue to next block if equal, exit to phi else.
667     Builder.CreateCondBr(IsEqual, NextCmpBlock, PhiBB);
668     Phi.addIncoming(ConstantInt::getFalse(Context), BB);
669     DTU.applyUpdates({{DominatorTree::Insert, BB, NextCmpBlock},
670                       {DominatorTree::Insert, BB, PhiBB}});
671   }
672   return BB;
673 }
674 
675 bool BCECmpChain::simplify(const TargetLibraryInfo &TLI, AliasAnalysis &AA,
676                            DomTreeUpdater &DTU) {
677   assert(atLeastOneMerged() && "simplifying trivial BCECmpChain");
678   LLVM_DEBUG(dbgs() << "Simplifying comparison chain starting at block "
679                     << EntryBlock_->getName() << "\n");
680 
681   // Effectively merge blocks. We go in the reverse direction from the phi block
682   // so that the next block is always available to branch to.
683   BasicBlock *InsertBefore = EntryBlock_;
684   BasicBlock *NextCmpBlock = Phi_.getParent();
685   for (const auto &Blocks : reverse(MergedBlocks_)) {
686     InsertBefore = NextCmpBlock = mergeComparisons(
687         Blocks, InsertBefore, NextCmpBlock, Phi_, TLI, AA, DTU);
688   }
689 
690   // Replace the original cmp chain with the new cmp chain by pointing all
691   // predecessors of EntryBlock_ to NextCmpBlock instead. This makes all cmp
692   // blocks in the old chain unreachable.
693   while (!pred_empty(EntryBlock_)) {
694     BasicBlock* const Pred = *pred_begin(EntryBlock_);
695     LLVM_DEBUG(dbgs() << "Updating jump into old chain from " << Pred->getName()
696                       << "\n");
697     Pred->getTerminator()->replaceUsesOfWith(EntryBlock_, NextCmpBlock);
698     DTU.applyUpdates({{DominatorTree::Delete, Pred, EntryBlock_},
699                       {DominatorTree::Insert, Pred, NextCmpBlock}});
700   }
701 
702   // If the old cmp chain was the function entry, we need to update the function
703   // entry.
704   const bool ChainEntryIsFnEntry = EntryBlock_->isEntryBlock();
705   if (ChainEntryIsFnEntry && DTU.hasDomTree()) {
706     LLVM_DEBUG(dbgs() << "Changing function entry from "
707                       << EntryBlock_->getName() << " to "
708                       << NextCmpBlock->getName() << "\n");
709     DTU.getDomTree().setNewRoot(NextCmpBlock);
710     DTU.applyUpdates({{DominatorTree::Delete, NextCmpBlock, EntryBlock_}});
711   }
712   EntryBlock_ = nullptr;
713 
714   // Delete merged blocks. This also removes incoming values in phi.
715   SmallVector<BasicBlock *, 16> DeadBlocks;
716   for (const auto &Blocks : MergedBlocks_) {
717     for (const BCECmpBlock &Block : Blocks) {
718       LLVM_DEBUG(dbgs() << "Deleting merged block " << Block.BB->getName()
719                         << "\n");
720       DeadBlocks.push_back(Block.BB);
721     }
722   }
723   DeleteDeadBlocks(DeadBlocks, &DTU);
724 
725   MergedBlocks_.clear();
726   return true;
727 }
728 
729 std::vector<BasicBlock *> getOrderedBlocks(PHINode &Phi,
730                                            BasicBlock *const LastBlock,
731                                            int NumBlocks) {
732   // Walk up from the last block to find other blocks.
733   std::vector<BasicBlock *> Blocks(NumBlocks);
734   assert(LastBlock && "invalid last block");
735   BasicBlock *CurBlock = LastBlock;
736   for (int BlockIndex = NumBlocks - 1; BlockIndex > 0; --BlockIndex) {
737     if (CurBlock->hasAddressTaken()) {
738       // Somebody is jumping to the block through an address, all bets are
739       // off.
740       LLVM_DEBUG(dbgs() << "skip: block " << BlockIndex
741                         << " has its address taken\n");
742       return {};
743     }
744     Blocks[BlockIndex] = CurBlock;
745     auto *SinglePredecessor = CurBlock->getSinglePredecessor();
746     if (!SinglePredecessor) {
747       // The block has two or more predecessors.
748       LLVM_DEBUG(dbgs() << "skip: block " << BlockIndex
749                         << " has two or more predecessors\n");
750       return {};
751     }
752     if (Phi.getBasicBlockIndex(SinglePredecessor) < 0) {
753       // The block does not link back to the phi.
754       LLVM_DEBUG(dbgs() << "skip: block " << BlockIndex
755                         << " does not link back to the phi\n");
756       return {};
757     }
758     CurBlock = SinglePredecessor;
759   }
760   Blocks[0] = CurBlock;
761   return Blocks;
762 }
763 
764 bool processPhi(PHINode &Phi, const TargetLibraryInfo &TLI, AliasAnalysis &AA,
765                 DomTreeUpdater &DTU) {
766   LLVM_DEBUG(dbgs() << "processPhi()\n");
767   if (Phi.getNumIncomingValues() <= 1) {
768     LLVM_DEBUG(dbgs() << "skip: only one incoming value in phi\n");
769     return false;
770   }
771   // We are looking for something that has the following structure:
772   //   bb1 --eq--> bb2 --eq--> bb3 --eq--> bb4 --+
773   //     \            \           \               \
774   //      ne           ne          ne              \
775   //       \            \           \               v
776   //        +------------+-----------+----------> bb_phi
777   //
778   //  - The last basic block (bb4 here) must branch unconditionally to bb_phi.
779   //    It's the only block that contributes a non-constant value to the Phi.
780   //  - All other blocks (b1, b2, b3) must have exactly two successors, one of
781   //    them being the phi block.
782   //  - All intermediate blocks (bb2, bb3) must have only one predecessor.
783   //  - Blocks cannot do other work besides the comparison, see doesOtherWork()
784 
785   // The blocks are not necessarily ordered in the phi, so we start from the
786   // last block and reconstruct the order.
787   BasicBlock *LastBlock = nullptr;
788   for (unsigned I = 0; I < Phi.getNumIncomingValues(); ++I) {
789     if (isa<ConstantInt>(Phi.getIncomingValue(I))) continue;
790     if (LastBlock) {
791       // There are several non-constant values.
792       LLVM_DEBUG(dbgs() << "skip: several non-constant values\n");
793       return false;
794     }
795     if (!isa<ICmpInst>(Phi.getIncomingValue(I)) ||
796         cast<ICmpInst>(Phi.getIncomingValue(I))->getParent() !=
797             Phi.getIncomingBlock(I)) {
798       // Non-constant incoming value is not from a cmp instruction or not
799       // produced by the last block. We could end up processing the value
800       // producing block more than once.
801       //
802       // This is an uncommon case, so we bail.
803       LLVM_DEBUG(
804           dbgs()
805           << "skip: non-constant value not from cmp or not from last block.\n");
806       return false;
807     }
808     LastBlock = Phi.getIncomingBlock(I);
809   }
810   if (!LastBlock) {
811     // There is no non-constant block.
812     LLVM_DEBUG(dbgs() << "skip: no non-constant block\n");
813     return false;
814   }
815   if (LastBlock->getSingleSuccessor() != Phi.getParent()) {
816     LLVM_DEBUG(dbgs() << "skip: last block non-phi successor\n");
817     return false;
818   }
819 
820   const auto Blocks =
821       getOrderedBlocks(Phi, LastBlock, Phi.getNumIncomingValues());
822   if (Blocks.empty()) return false;
823   BCECmpChain CmpChain(Blocks, Phi, AA);
824 
825   if (!CmpChain.atLeastOneMerged()) {
826     LLVM_DEBUG(dbgs() << "skip: nothing merged\n");
827     return false;
828   }
829 
830   return CmpChain.simplify(TLI, AA, DTU);
831 }
832 
833 static bool runImpl(Function &F, const TargetLibraryInfo &TLI,
834                     const TargetTransformInfo &TTI, AliasAnalysis &AA,
835                     DominatorTree *DT) {
836   LLVM_DEBUG(dbgs() << "MergeICmpsLegacyPass: " << F.getName() << "\n");
837 
838   // We only try merging comparisons if the target wants to expand memcmp later.
839   // The rationale is to avoid turning small chains into memcmp calls.
840   if (!TTI.enableMemCmpExpansion(F.hasOptSize(), true))
841     return false;
842 
843   // If we don't have memcmp avaiable we can't emit calls to it.
844   if (!TLI.has(LibFunc_memcmp))
845     return false;
846 
847   DomTreeUpdater DTU(DT, /*PostDominatorTree*/ nullptr,
848                      DomTreeUpdater::UpdateStrategy::Eager);
849 
850   bool MadeChange = false;
851 
852   for (BasicBlock &BB : llvm::drop_begin(F)) {
853     // A Phi operation is always first in a basic block.
854     if (auto *const Phi = dyn_cast<PHINode>(&*BB.begin()))
855       MadeChange |= processPhi(*Phi, TLI, AA, DTU);
856   }
857 
858   return MadeChange;
859 }
860 
861 class MergeICmpsLegacyPass : public FunctionPass {
862 public:
863   static char ID;
864 
865   MergeICmpsLegacyPass() : FunctionPass(ID) {
866     initializeMergeICmpsLegacyPassPass(*PassRegistry::getPassRegistry());
867   }
868 
869   bool runOnFunction(Function &F) override {
870     if (skipFunction(F)) return false;
871     const auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
872     const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
873     // MergeICmps does not need the DominatorTree, but we update it if it's
874     // already available.
875     auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
876     auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
877     return runImpl(F, TLI, TTI, AA, DTWP ? &DTWP->getDomTree() : nullptr);
878   }
879 
880  private:
881   void getAnalysisUsage(AnalysisUsage &AU) const override {
882     AU.addRequired<TargetLibraryInfoWrapperPass>();
883     AU.addRequired<TargetTransformInfoWrapperPass>();
884     AU.addRequired<AAResultsWrapperPass>();
885     AU.addPreserved<GlobalsAAWrapperPass>();
886     AU.addPreserved<DominatorTreeWrapperPass>();
887   }
888 };
889 
890 } // namespace
891 
892 char MergeICmpsLegacyPass::ID = 0;
893 INITIALIZE_PASS_BEGIN(MergeICmpsLegacyPass, "mergeicmps",
894                       "Merge contiguous icmps into a memcmp", false, false)
895 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
896 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
897 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
898 INITIALIZE_PASS_END(MergeICmpsLegacyPass, "mergeicmps",
899                     "Merge contiguous icmps into a memcmp", false, false)
900 
901 Pass *llvm::createMergeICmpsLegacyPass() { return new MergeICmpsLegacyPass(); }
902 
903 PreservedAnalyses MergeICmpsPass::run(Function &F,
904                                       FunctionAnalysisManager &AM) {
905   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
906   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
907   auto &AA = AM.getResult<AAManager>(F);
908   auto *DT = AM.getCachedResult<DominatorTreeAnalysis>(F);
909   const bool MadeChanges = runImpl(F, TLI, TTI, AA, DT);
910   if (!MadeChanges)
911     return PreservedAnalyses::all();
912   PreservedAnalyses PA;
913   PA.preserve<DominatorTreeAnalysis>();
914   return PA;
915 }
916