1 //===- MergeICmps.cpp - Optimize chains of integer comparisons ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass turns chains of integer comparisons into memcmp (the memcmp is
11 // later typically inlined as a chain of efficient hardware comparisons). This
12 // typically benefits c++ member or nonmember operator==().
14 // The basic idea is to replace a larger chain of integer comparisons loaded
15 // from contiguous memory locations into a smaller chain of such integer
16 // comparisons. Benefits are double:
17 // - There are less jumps, and therefore less opportunities for mispredictions
18 // and I-cache misses.
19 // - Code size is smaller, both because jumps are removed and because the
20 // encoding of a 2*n byte compare is smaller than that of two n-byte
23 //===----------------------------------------------------------------------===//
29 #include "llvm/Analysis/Loads.h"
30 #include "llvm/Analysis/TargetLibraryInfo.h"
31 #include "llvm/Analysis/TargetTransformInfo.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/IRBuilder.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Transforms/Scalar.h"
36 #include "llvm/Transforms/Utils/BuildLibCalls.h"
42 #define DEBUG_TYPE "mergeicmps"
46 BCEAtom() : GEP(nullptr), LoadI(nullptr), Offset() {}
48 const Value *Base() const { return GEP ? GEP->getPointerOperand() : nullptr; }
50 bool operator<(const BCEAtom &O) const {
51 assert(Base() && "invalid atom");
52 assert(O.Base() && "invalid atom");
53 // Just ordering by (Base(), Offset) is sufficient. However because this
54 // means that the ordering will depend on the addresses of the base
55 // values, which are not reproducible from run to run. To guarantee
56 // stability, we use the names of the values if they exist; we sort by:
57 // (Base.getName(), Base(), Offset).
58 const int NameCmp = Base()->getName().compare(O.Base()->getName());
60 if (Base() == O.Base()) {
61 return Offset.slt(O.Offset);
63 return Base() < O.Base();
68 GetElementPtrInst *GEP;
73 // If this value is a load from a constant offset w.r.t. a base address, and
74 // there are no othe rusers of the load or address, returns the base address and
76 BCEAtom visitICmpLoadOperand(Value *const Val) {
78 if (auto *const LoadI = dyn_cast<LoadInst>(Val)) {
79 DEBUG(dbgs() << "load\n");
80 if (LoadI->isUsedOutsideOfBlock(LoadI->getParent())) {
81 DEBUG(dbgs() << "used outside of block\n");
84 if (LoadI->isVolatile()) {
85 DEBUG(dbgs() << "volatile\n");
88 Value *const Addr = LoadI->getOperand(0);
89 if (auto *const GEP = dyn_cast<GetElementPtrInst>(Addr)) {
90 DEBUG(dbgs() << "GEP\n");
91 if (LoadI->isUsedOutsideOfBlock(LoadI->getParent())) {
92 DEBUG(dbgs() << "used outside of block\n");
95 const auto &DL = GEP->getModule()->getDataLayout();
96 if (!isDereferenceablePointer(GEP, DL)) {
97 DEBUG(dbgs() << "not dereferenceable\n");
98 // We need to make sure that we can do comparison in any order, so we
99 // require memory to be unconditionnally dereferencable.
102 Result.Offset = APInt(DL.getPointerTypeSizeInBits(GEP->getType()), 0);
103 if (GEP->accumulateConstantOffset(DL, Result.Offset)) {
105 Result.LoadI = LoadI;
112 // A basic block with a comparison between two BCE atoms.
113 // Note: the terminology is misleading: the comparison is symmetric, so there
114 // is no real {l/r}hs. What we want though is to have the same base on the
115 // left (resp. right), so that we can detect consecutive loads. To ensure this
116 // we put the smallest atom on the left.
121 BCECmpBlock(BCEAtom L, BCEAtom R, int SizeBits)
122 : Lhs_(L), Rhs_(R), SizeBits_(SizeBits) {
123 if (Rhs_ < Lhs_) std::swap(Rhs_, Lhs_);
126 bool IsValid() const {
127 return Lhs_.Base() != nullptr && Rhs_.Base() != nullptr;
130 // Assert the the block is consistent: If valid, it should also have
131 // non-null members besides Lhs_ and Rhs_.
132 void AssertConsistent() const {
140 const BCEAtom &Lhs() const { return Lhs_; }
141 const BCEAtom &Rhs() const { return Rhs_; }
142 int SizeBits() const { return SizeBits_; }
144 // Returns true if the block does other works besides comparison.
145 bool doesOtherWork() const;
147 // The basic block where this comparison happens.
148 BasicBlock *BB = nullptr;
149 // The ICMP for this comparison.
150 ICmpInst *CmpI = nullptr;
151 // The terminating branch.
152 BranchInst *BranchI = nullptr;
160 bool BCECmpBlock::doesOtherWork() const {
162 // TODO(courbet): Can we allow some other things ? This is very conservative.
163 // We might be able to get away with anything does does not have any side
164 // effects outside of the basic block.
165 // Note: The GEPs and/or loads are not necessarily in the same block.
166 for (const Instruction &Inst : *BB) {
167 if (const auto *const GEP = dyn_cast<GetElementPtrInst>(&Inst)) {
168 if (!(Lhs_.GEP == GEP || Rhs_.GEP == GEP)) return true;
169 } else if (const auto *const L = dyn_cast<LoadInst>(&Inst)) {
170 if (!(Lhs_.LoadI == L || Rhs_.LoadI == L)) return true;
171 } else if (const auto *const C = dyn_cast<ICmpInst>(&Inst)) {
172 if (C != CmpI) return true;
173 } else if (const auto *const Br = dyn_cast<BranchInst>(&Inst)) {
174 if (Br != BranchI) return true;
182 // Visit the given comparison. If this is a comparison between two valid
183 // BCE atoms, returns the comparison.
184 BCECmpBlock visitICmp(const ICmpInst *const CmpI,
185 const ICmpInst::Predicate ExpectedPredicate) {
186 if (CmpI->getPredicate() == ExpectedPredicate) {
187 DEBUG(dbgs() << "cmp "
188 << (ExpectedPredicate == ICmpInst::ICMP_EQ ? "eq" : "ne")
190 auto Lhs = visitICmpLoadOperand(CmpI->getOperand(0));
191 if (!Lhs.Base()) return {};
192 auto Rhs = visitICmpLoadOperand(CmpI->getOperand(1));
193 if (!Rhs.Base()) return {};
194 return BCECmpBlock(std::move(Lhs), std::move(Rhs),
195 CmpI->getOperand(0)->getType()->getScalarSizeInBits());
200 // Visit the given comparison block. If this is a comparison between two valid
201 // BCE atoms, returns the comparison.
202 BCECmpBlock visitCmpBlock(Value *const Val, BasicBlock *const Block,
203 const BasicBlock *const PhiBlock) {
204 if (Block->empty()) return {};
205 auto *const BranchI = dyn_cast<BranchInst>(Block->getTerminator());
206 if (!BranchI) return {};
207 DEBUG(dbgs() << "branch\n");
208 if (BranchI->isUnconditional()) {
209 // In this case, we expect an incoming value which is the result of the
210 // comparison. This is the last link in the chain of comparisons (note
211 // that this does not mean that this is the last incoming value, blocks
212 // can be reordered).
213 auto *const CmpI = dyn_cast<ICmpInst>(Val);
214 if (!CmpI) return {};
215 DEBUG(dbgs() << "icmp\n");
216 auto Result = visitICmp(CmpI, ICmpInst::ICMP_EQ);
218 Result.BranchI = BranchI;
221 // In this case, we expect a constant incoming value (the comparison is
223 const auto *const Const = dyn_cast<ConstantInt>(Val);
224 DEBUG(dbgs() << "const\n");
225 if (!Const->isZero()) return {};
226 DEBUG(dbgs() << "false\n");
227 auto *const CmpI = dyn_cast<ICmpInst>(BranchI->getCondition());
228 if (!CmpI) return {};
229 DEBUG(dbgs() << "icmp\n");
230 assert(BranchI->getNumSuccessors() == 2 && "expecting a cond branch");
231 BasicBlock *const FalseBlock = BranchI->getSuccessor(1);
232 auto Result = visitICmp(
233 CmpI, FalseBlock == PhiBlock ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE);
235 Result.BranchI = BranchI;
241 // A chain of comparisons.
244 BCECmpChain(const std::vector<BasicBlock *> &Blocks, PHINode &Phi);
246 int size() const { return Comparisons_.size(); }
248 #ifdef MERGEICMPS_DOT_ON
250 #endif // MERGEICMPS_DOT_ON
252 bool simplify(const TargetLibraryInfo *const TLI);
255 static bool IsContiguous(const BCECmpBlock &First,
256 const BCECmpBlock &Second) {
257 return First.Lhs().Base() == Second.Lhs().Base() &&
258 First.Rhs().Base() == Second.Rhs().Base() &&
259 First.Lhs().Offset + First.SizeBits() / 8 == Second.Lhs().Offset &&
260 First.Rhs().Offset + First.SizeBits() / 8 == Second.Rhs().Offset;
263 // Merges the given comparison blocks into one memcmp block and update
264 // branches. Comparisons are assumed to be continguous. If NextBBInChain is
265 // null, the merged block will link to the phi block.
266 static void mergeComparisons(ArrayRef<BCECmpBlock> Comparisons,
267 BasicBlock *const NextBBInChain, PHINode &Phi,
268 const TargetLibraryInfo *const TLI);
271 std::vector<BCECmpBlock> Comparisons_;
272 // The original entry block (before sorting);
273 BasicBlock *EntryBlock_;
276 BCECmpChain::BCECmpChain(const std::vector<BasicBlock *> &Blocks, PHINode &Phi)
278 // Now look inside blocks to check for BCE comparisons.
279 std::vector<BCECmpBlock> Comparisons;
280 for (BasicBlock *Block : Blocks) {
281 BCECmpBlock Comparison = visitCmpBlock(Phi.getIncomingValueForBlock(Block),
282 Block, Phi.getParent());
283 Comparison.BB = Block;
284 if (!Comparison.IsValid()) {
285 DEBUG(dbgs() << "skip: not a valid BCECmpBlock\n");
288 if (Comparison.doesOtherWork()) {
289 DEBUG(dbgs() << "block does extra work besides compare\n");
290 if (Comparisons.empty()) { // First block.
291 // TODO(courbet): The first block can do other things, and we should
292 // split them apart in a separate block before the comparison chain.
293 // Right now we just discard it and make the chain shorter.
295 << "ignoring first block that does extra work besides compare\n");
298 // TODO(courbet): Right now we abort the whole chain. We could be
299 // merging only the blocks that don't do other work and resume the
300 // chain from there. For example:
301 // if (a[0] == b[0]) { // bb1
302 // if (a[1] == b[1]) { // bb2
303 // some_value = 3; //bb3
304 // if (a[2] == b[2]) { //bb3
305 // do a ton of stuff //bb4
312 // bb1 --eq--> bb2 --eq--> bb3* -eq--> bb4 --+
316 // +------------+-----------+----------> bb_phi
318 // We can only merge the first two comparisons, because bb3* does
319 // "other work" (setting some_value to 3).
320 // We could still merge bb1 and bb2 though.
323 DEBUG(dbgs() << "*Found cmp of " << Comparison.SizeBits()
324 << " bits between " << Comparison.Lhs().Base() << " + "
325 << Comparison.Lhs().Offset << " and "
326 << Comparison.Rhs().Base() << " + " << Comparison.Rhs().Offset
328 DEBUG(dbgs() << "\n");
329 Comparisons.push_back(Comparison);
331 EntryBlock_ = Comparisons[0].BB;
332 Comparisons_ = std::move(Comparisons);
333 #ifdef MERGEICMPS_DOT_ON
334 errs() << "BEFORE REORDERING:\n\n";
336 #endif // MERGEICMPS_DOT_ON
337 // Reorder blocks by LHS. We can do that without changing the
338 // semantics because we are only accessing dereferencable memory.
339 std::sort(Comparisons_.begin(), Comparisons_.end(),
340 [](const BCECmpBlock &a, const BCECmpBlock &b) {
341 return a.Lhs() < b.Lhs();
343 #ifdef MERGEICMPS_DOT_ON
344 errs() << "AFTER REORDERING:\n\n";
346 #endif // MERGEICMPS_DOT_ON
349 #ifdef MERGEICMPS_DOT_ON
350 void BCECmpChain::dump() const {
351 errs() << "digraph dag {\n";
352 errs() << " graph [bgcolor=transparent];\n";
353 errs() << " node [color=black,style=filled,fillcolor=lightyellow];\n";
354 errs() << " edge [color=black];\n";
355 for (size_t I = 0; I < Comparisons_.size(); ++I) {
356 const auto &Comparison = Comparisons_[I];
357 errs() << " \"" << I << "\" [label=\"%"
358 << Comparison.Lhs().Base()->getName() << " + "
359 << Comparison.Lhs().Offset << " == %"
360 << Comparison.Rhs().Base()->getName() << " + "
361 << Comparison.Rhs().Offset << " (" << (Comparison.SizeBits() / 8)
363 const Value *const Val = Phi_.getIncomingValueForBlock(Comparison.BB);
364 if (I > 0) errs() << " \"" << (I - 1) << "\" -> \"" << I << "\";\n";
365 errs() << " \"" << I << "\" -> \"Phi\" [label=\"" << *Val << "\"];\n";
367 errs() << " \"Phi\" [label=\"Phi\"];\n";
370 #endif // MERGEICMPS_DOT_ON
372 bool BCECmpChain::simplify(const TargetLibraryInfo *const TLI) {
373 // First pass to check if there is at least one merge. If not, we don't do
374 // anything and we keep analysis passes intact.
376 bool AtLeastOneMerged = false;
377 for (size_t I = 1; I < Comparisons_.size(); ++I) {
378 if (IsContiguous(Comparisons_[I - 1], Comparisons_[I])) {
379 AtLeastOneMerged = true;
383 if (!AtLeastOneMerged) return false;
386 // Remove phi references to comparison blocks, they will be rebuilt as we
388 for (const auto &Comparison : Comparisons_) {
389 Phi_.removeIncomingValue(Comparison.BB, false);
392 // Point the predecessors of the chain to the first comparison block (which is
393 // the new entry point).
394 if (EntryBlock_ != Comparisons_[0].BB)
395 EntryBlock_->replaceAllUsesWith(Comparisons_[0].BB);
397 // Effectively merge blocks.
399 for (size_t I = 1; I < Comparisons_.size(); ++I) {
400 if (IsContiguous(Comparisons_[I - 1], Comparisons_[I])) {
403 // Merge all previous comparisons and start a new merge block.
405 makeArrayRef(Comparisons_).slice(I - NumMerged, NumMerged),
406 Comparisons_[I].BB, Phi_, TLI);
410 mergeComparisons(makeArrayRef(Comparisons_)
411 .slice(Comparisons_.size() - NumMerged, NumMerged),
417 void BCECmpChain::mergeComparisons(ArrayRef<BCECmpBlock> Comparisons,
418 BasicBlock *const NextBBInChain,
420 const TargetLibraryInfo *const TLI) {
421 assert(!Comparisons.empty());
422 const auto &FirstComparison = *Comparisons.begin();
423 BasicBlock *const BB = FirstComparison.BB;
424 LLVMContext &Context = BB->getContext();
426 if (Comparisons.size() >= 2) {
427 DEBUG(dbgs() << "Merging " << Comparisons.size() << " comparisons\n");
428 const auto TotalSize =
429 std::accumulate(Comparisons.begin(), Comparisons.end(), 0,
430 [](int Size, const BCECmpBlock &C) {
431 return Size + C.SizeBits();
435 // Incoming edges do not need to be updated, and both GEPs are already
436 // computing the right address, we just need to:
437 // - replace the two loads and the icmp with the memcmp
438 // - update the branch
439 // - update the incoming values in the phi.
440 FirstComparison.BranchI->eraseFromParent();
441 FirstComparison.CmpI->eraseFromParent();
442 FirstComparison.Lhs().LoadI->eraseFromParent();
443 FirstComparison.Rhs().LoadI->eraseFromParent();
445 IRBuilder<> Builder(BB);
446 const auto &DL = Phi.getModule()->getDataLayout();
447 Value *const MemCmpCall = emitMemCmp(
448 FirstComparison.Lhs().GEP, FirstComparison.Rhs().GEP, ConstantInt::get(DL.getIntPtrType(Context), TotalSize),
450 Value *const MemCmpIsZero = Builder.CreateICmpEQ(
451 MemCmpCall, ConstantInt::get(Type::getInt32Ty(Context), 0));
453 // Add a branch to the next basic block in the chain.
455 Builder.CreateCondBr(MemCmpIsZero, NextBBInChain, Phi.getParent());
456 Phi.addIncoming(ConstantInt::getFalse(Context), BB);
458 Builder.CreateBr(Phi.getParent());
459 Phi.addIncoming(MemCmpIsZero, BB);
462 // Delete merged blocks.
463 for (size_t I = 1; I < Comparisons.size(); ++I) {
464 BasicBlock *CBB = Comparisons[I].BB;
465 CBB->replaceAllUsesWith(BB);
466 CBB->eraseFromParent();
469 assert(Comparisons.size() == 1);
470 // There are no blocks to merge, but we still need to update the branches.
471 DEBUG(dbgs() << "Only one comparison, updating branches\n");
473 if (FirstComparison.BranchI->isConditional()) {
474 DEBUG(dbgs() << "conditional -> conditional\n");
475 // Just update the "true" target, the "false" target should already be
477 assert(FirstComparison.BranchI->getSuccessor(1) == Phi.getParent());
478 FirstComparison.BranchI->setSuccessor(0, NextBBInChain);
479 Phi.addIncoming(ConstantInt::getFalse(Context), BB);
481 DEBUG(dbgs() << "unconditional -> conditional\n");
482 // Replace the unconditional branch by a conditional one.
483 FirstComparison.BranchI->eraseFromParent();
484 IRBuilder<> Builder(BB);
485 Builder.CreateCondBr(FirstComparison.CmpI, NextBBInChain,
487 Phi.addIncoming(FirstComparison.CmpI, BB);
490 if (FirstComparison.BranchI->isConditional()) {
491 DEBUG(dbgs() << "conditional -> unconditional\n");
492 // Replace the conditional branch by an unconditional one.
493 FirstComparison.BranchI->eraseFromParent();
494 IRBuilder<> Builder(BB);
495 Builder.CreateBr(Phi.getParent());
496 Phi.addIncoming(FirstComparison.CmpI, BB);
498 DEBUG(dbgs() << "unconditional -> unconditional\n");
499 Phi.addIncoming(FirstComparison.CmpI, BB);
505 std::vector<BasicBlock *> getOrderedBlocks(PHINode &Phi,
506 BasicBlock *const LastBlock,
508 // Walk up from the last block to find other blocks.
509 std::vector<BasicBlock *> Blocks(NumBlocks);
510 BasicBlock *CurBlock = LastBlock;
511 for (int BlockIndex = NumBlocks - 1; BlockIndex > 0; --BlockIndex) {
512 if (CurBlock->hasAddressTaken()) {
513 // Somebody is jumping to the block through an address, all bets are
515 DEBUG(dbgs() << "skip: block " << BlockIndex
516 << " has its address taken\n");
519 Blocks[BlockIndex] = CurBlock;
520 auto *SinglePredecessor = CurBlock->getSinglePredecessor();
521 if (!SinglePredecessor) {
522 // The block has two or more predecessors.
523 DEBUG(dbgs() << "skip: block " << BlockIndex
524 << " has two or more predecessors\n");
527 if (Phi.getBasicBlockIndex(SinglePredecessor) < 0) {
528 // The block does not link back to the phi.
529 DEBUG(dbgs() << "skip: block " << BlockIndex
530 << " does not link back to the phi\n");
533 CurBlock = SinglePredecessor;
535 Blocks[0] = CurBlock;
539 bool processPhi(PHINode &Phi, const TargetLibraryInfo *const TLI) {
540 DEBUG(dbgs() << "processPhi()\n");
541 if (Phi.getNumIncomingValues() <= 1) {
542 DEBUG(dbgs() << "skip: only one incoming value in phi\n");
545 // We are looking for something that has the following structure:
546 // bb1 --eq--> bb2 --eq--> bb3 --eq--> bb4 --+
550 // +------------+-----------+----------> bb_phi
552 // - The last basic block (bb4 here) must branch unconditionally to bb_phi.
553 // It's the only block that contributes a non-constant value to the Phi.
554 // - All other blocks (b1, b2, b3) must have exactly two successors, one of
555 // them being the the phi block.
556 // - All intermediate blocks (bb2, bb3) must have only one predecessor.
557 // - Blocks cannot do other work besides the comparison, see doesOtherWork()
559 // The blocks are not necessarily ordered in the phi, so we start from the
560 // last block and reconstruct the order.
561 BasicBlock *LastBlock = nullptr;
562 for (unsigned I = 0; I < Phi.getNumIncomingValues(); ++I) {
563 if (isa<ConstantInt>(Phi.getIncomingValue(I))) continue;
565 // There are several non-constant values.
566 DEBUG(dbgs() << "skip: several non-constant values\n");
569 LastBlock = Phi.getIncomingBlock(I);
572 // There is no non-constant block.
573 DEBUG(dbgs() << "skip: no non-constant block\n");
576 if (LastBlock->getSingleSuccessor() != Phi.getParent()) {
577 DEBUG(dbgs() << "skip: last block non-phi successor\n");
582 getOrderedBlocks(Phi, LastBlock, Phi.getNumIncomingValues());
583 if (Blocks.empty()) return false;
584 BCECmpChain CmpChain(Blocks, Phi);
586 if (CmpChain.size() < 2) {
587 DEBUG(dbgs() << "skip: only one compare block\n");
591 return CmpChain.simplify(TLI);
594 class MergeICmps : public FunctionPass {
598 MergeICmps() : FunctionPass(ID) {
599 initializeMergeICmpsPass(*PassRegistry::getPassRegistry());
602 bool runOnFunction(Function &F) override {
603 if (skipFunction(F)) return false;
604 const auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
605 const auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
606 auto PA = runImpl(F, &TLI, &TTI);
607 return !PA.areAllPreserved();
611 void getAnalysisUsage(AnalysisUsage &AU) const override {
612 AU.addRequired<TargetLibraryInfoWrapperPass>();
613 AU.addRequired<TargetTransformInfoWrapperPass>();
616 PreservedAnalyses runImpl(Function &F, const TargetLibraryInfo *TLI,
617 const TargetTransformInfo *TTI);
620 PreservedAnalyses MergeICmps::runImpl(Function &F, const TargetLibraryInfo *TLI,
621 const TargetTransformInfo *TTI) {
622 DEBUG(dbgs() << "MergeICmpsPass: " << F.getName() << "\n");
624 // We only try merging comparisons if the target wants to expand memcmp later.
625 // The rationale is to avoid turning small chains into memcmp calls.
626 if (!TTI->enableMemCmpExpansion(true)) return PreservedAnalyses::all();
628 bool MadeChange = false;
630 for (auto BBIt = ++F.begin(); BBIt != F.end(); ++BBIt) {
631 // A Phi operation is always first in a basic block.
632 if (auto *const Phi = dyn_cast<PHINode>(&*BBIt->begin()))
633 MadeChange |= processPhi(*Phi, TLI);
636 if (MadeChange) return PreservedAnalyses::none();
637 return PreservedAnalyses::all();
642 char MergeICmps::ID = 0;
643 INITIALIZE_PASS_BEGIN(MergeICmps, "mergeicmps",
644 "Merge contiguous icmps into a memcmp", false, false)
645 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
646 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
647 INITIALIZE_PASS_END(MergeICmps, "mergeicmps",
648 "Merge contiguous icmps into a memcmp", false, false)
650 Pass *llvm::createMergeICmpsPass() { return new MergeICmps(); }