//===- LoopIdiomRecognize.cpp - Loop idiom recognition --------------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This pass implements an idiom recognizer that transforms simple loops into a // non-loop form. In cases that this kicks in, it can be a significant // performance win. // // If compiling for code size we avoid idiom recognition if the resulting // code could be larger than the code for the original loop. One way this could // happen is if the loop is not removable after idiom recognition due to the // presence of non-idiom instructions. The initial implementation of the // heuristics applies to idioms in multi-block loops. // //===----------------------------------------------------------------------===// // // TODO List: // // Future loop memory idioms to recognize: // memcmp, strlen, etc. // Future floating point idioms to recognize in -ffast-math mode: // fpowi // Future integer operation idioms to recognize: // ctpop // // Beware that isel's default lowering for ctpop is highly inefficient for // i64 and larger types when i64 is legal and the value has few bits set. It // would be good to enhance isel to emit a loop for ctpop in this case. // // This could recognize common matrix multiplies and dot product idioms and // replace them with calls to BLAS (if linked in??). // //===----------------------------------------------------------------------===// #include "llvm/Transforms/Scalar/LoopIdiomRecognize.h" #include "llvm/ADT/APInt.h" #include "llvm/ADT/ArrayRef.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/MapVector.h" #include "llvm/ADT/SetVector.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/StringRef.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/Analysis/CmpInstAnalysis.h" #include "llvm/Analysis/LoopAccessAnalysis.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/LoopPass.h" #include "llvm/Analysis/MemoryLocation.h" #include "llvm/Analysis/MemorySSA.h" #include "llvm/Analysis/MemorySSAUpdater.h" #include "llvm/Analysis/MustExecute.h" #include "llvm/Analysis/OptimizationRemarkEmitter.h" #include "llvm/Analysis/ScalarEvolution.h" #include "llvm/Analysis/ScalarEvolutionExpressions.h" #include "llvm/Analysis/TargetLibraryInfo.h" #include "llvm/Analysis/TargetTransformInfo.h" #include "llvm/Analysis/ValueTracking.h" #include "llvm/IR/BasicBlock.h" #include "llvm/IR/Constant.h" #include "llvm/IR/Constants.h" #include "llvm/IR/DataLayout.h" #include "llvm/IR/DebugLoc.h" #include "llvm/IR/DerivedTypes.h" #include "llvm/IR/Dominators.h" #include "llvm/IR/GlobalValue.h" #include "llvm/IR/GlobalVariable.h" #include "llvm/IR/IRBuilder.h" #include "llvm/IR/InstrTypes.h" #include "llvm/IR/Instruction.h" #include "llvm/IR/Instructions.h" #include "llvm/IR/IntrinsicInst.h" #include "llvm/IR/Intrinsics.h" #include "llvm/IR/LLVMContext.h" #include "llvm/IR/Module.h" #include "llvm/IR/PassManager.h" #include "llvm/IR/PatternMatch.h" #include "llvm/IR/Type.h" #include "llvm/IR/User.h" #include "llvm/IR/Value.h" #include "llvm/IR/ValueHandle.h" #include "llvm/Support/Casting.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/InstructionCost.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Transforms/Utils/BuildLibCalls.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Utils/LoopUtils.h" #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" #include #include #include #include #include using namespace llvm; #define DEBUG_TYPE "loop-idiom" STATISTIC(NumMemSet, "Number of memset's formed from loop stores"); STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores"); STATISTIC(NumMemMove, "Number of memmove's formed from loop load+stores"); STATISTIC( NumShiftUntilBitTest, "Number of uncountable loops recognized as 'shift until bitttest' idiom"); STATISTIC(NumShiftUntilZero, "Number of uncountable loops recognized as 'shift until zero' idiom"); bool DisableLIRP::All; static cl::opt DisableLIRPAll("disable-" DEBUG_TYPE "-all", cl::desc("Options to disable Loop Idiom Recognize Pass."), cl::location(DisableLIRP::All), cl::init(false), cl::ReallyHidden); bool DisableLIRP::Memset; static cl::opt DisableLIRPMemset("disable-" DEBUG_TYPE "-memset", cl::desc("Proceed with loop idiom recognize pass, but do " "not convert loop(s) to memset."), cl::location(DisableLIRP::Memset), cl::init(false), cl::ReallyHidden); bool DisableLIRP::Memcpy; static cl::opt DisableLIRPMemcpy("disable-" DEBUG_TYPE "-memcpy", cl::desc("Proceed with loop idiom recognize pass, but do " "not convert loop(s) to memcpy."), cl::location(DisableLIRP::Memcpy), cl::init(false), cl::ReallyHidden); static cl::opt UseLIRCodeSizeHeurs( "use-lir-code-size-heurs", cl::desc("Use loop idiom recognition code size heuristics when compiling" "with -Os/-Oz"), cl::init(true), cl::Hidden); namespace { class LoopIdiomRecognize { Loop *CurLoop = nullptr; AliasAnalysis *AA; DominatorTree *DT; LoopInfo *LI; ScalarEvolution *SE; TargetLibraryInfo *TLI; const TargetTransformInfo *TTI; const DataLayout *DL; OptimizationRemarkEmitter &ORE; bool ApplyCodeSizeHeuristics; std::unique_ptr MSSAU; public: explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT, LoopInfo *LI, ScalarEvolution *SE, TargetLibraryInfo *TLI, const TargetTransformInfo *TTI, MemorySSA *MSSA, const DataLayout *DL, OptimizationRemarkEmitter &ORE) : AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI), DL(DL), ORE(ORE) { if (MSSA) MSSAU = std::make_unique(MSSA); } bool runOnLoop(Loop *L); private: using StoreList = SmallVector; using StoreListMap = MapVector; StoreListMap StoreRefsForMemset; StoreListMap StoreRefsForMemsetPattern; StoreList StoreRefsForMemcpy; bool HasMemset; bool HasMemsetPattern; bool HasMemcpy; /// Return code for isLegalStore() enum LegalStoreKind { None = 0, Memset, MemsetPattern, Memcpy, UnorderedAtomicMemcpy, DontUse // Dummy retval never to be used. Allows catching errors in retval // handling. }; /// \name Countable Loop Idiom Handling /// @{ bool runOnCountableLoop(); bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount, SmallVectorImpl &ExitBlocks); void collectStores(BasicBlock *BB); LegalStoreKind isLegalStore(StoreInst *SI); enum class ForMemset { No, Yes }; bool processLoopStores(SmallVectorImpl &SL, const SCEV *BECount, ForMemset For); template bool processLoopMemIntrinsic( BasicBlock *BB, bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *), const SCEV *BECount); bool processLoopMemCpy(MemCpyInst *MCI, const SCEV *BECount); bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount); bool processLoopStridedStore(Value *DestPtr, const SCEV *StoreSizeSCEV, MaybeAlign StoreAlignment, Value *StoredVal, Instruction *TheStore, SmallPtrSetImpl &Stores, const SCEVAddRecExpr *Ev, const SCEV *BECount, bool IsNegStride, bool IsLoopMemset = false); bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount); bool processLoopStoreOfLoopLoad(Value *DestPtr, Value *SourcePtr, const SCEV *StoreSize, MaybeAlign StoreAlign, MaybeAlign LoadAlign, Instruction *TheStore, Instruction *TheLoad, const SCEVAddRecExpr *StoreEv, const SCEVAddRecExpr *LoadEv, const SCEV *BECount); bool avoidLIRForMultiBlockLoop(bool IsMemset = false, bool IsLoopMemset = false); /// @} /// \name Noncountable Loop Idiom Handling /// @{ bool runOnNoncountableLoop(); bool recognizePopcount(); void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst, PHINode *CntPhi, Value *Var); bool recognizeAndInsertFFS(); /// Find First Set: ctlz or cttz void transformLoopToCountable(Intrinsic::ID IntrinID, BasicBlock *PreCondBB, Instruction *CntInst, PHINode *CntPhi, Value *Var, Instruction *DefX, const DebugLoc &DL, bool ZeroCheck, bool IsCntPhiUsedOutsideLoop); bool recognizeShiftUntilBitTest(); bool recognizeShiftUntilZero(); /// @} }; } // end anonymous namespace PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM, LoopStandardAnalysisResults &AR, LPMUpdater &) { if (DisableLIRP::All) return PreservedAnalyses::all(); const auto *DL = &L.getHeader()->getModule()->getDataLayout(); // For the new PM, we also can't use OptimizationRemarkEmitter as an analysis // pass. Function analyses need to be preserved across loop transformations // but ORE cannot be preserved (see comment before the pass definition). OptimizationRemarkEmitter ORE(L.getHeader()->getParent()); LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI, AR.MSSA, DL, ORE); if (!LIR.runOnLoop(&L)) return PreservedAnalyses::all(); auto PA = getLoopPassPreservedAnalyses(); if (AR.MSSA) PA.preserve(); return PA; } static void deleteDeadInstruction(Instruction *I) { I->replaceAllUsesWith(PoisonValue::get(I->getType())); I->eraseFromParent(); } //===----------------------------------------------------------------------===// // // Implementation of LoopIdiomRecognize // //===----------------------------------------------------------------------===// bool LoopIdiomRecognize::runOnLoop(Loop *L) { CurLoop = L; // If the loop could not be converted to canonical form, it must have an // indirectbr in it, just give up. if (!L->getLoopPreheader()) return false; // Disable loop idiom recognition if the function's name is a common idiom. StringRef Name = L->getHeader()->getParent()->getName(); if (Name == "memset" || Name == "memcpy") return false; // Determine if code size heuristics need to be applied. ApplyCodeSizeHeuristics = L->getHeader()->getParent()->hasOptSize() && UseLIRCodeSizeHeurs; HasMemset = TLI->has(LibFunc_memset); HasMemsetPattern = TLI->has(LibFunc_memset_pattern16); HasMemcpy = TLI->has(LibFunc_memcpy); if (HasMemset || HasMemsetPattern || HasMemcpy) if (SE->hasLoopInvariantBackedgeTakenCount(L)) return runOnCountableLoop(); return runOnNoncountableLoop(); } bool LoopIdiomRecognize::runOnCountableLoop() { const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop); assert(!isa(BECount) && "runOnCountableLoop() called on a loop without a predictable" "backedge-taken count"); // If this loop executes exactly one time, then it should be peeled, not // optimized by this pass. if (const SCEVConstant *BECst = dyn_cast(BECount)) if (BECst->getAPInt() == 0) return false; SmallVector ExitBlocks; CurLoop->getUniqueExitBlocks(ExitBlocks); LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F[" << CurLoop->getHeader()->getParent()->getName() << "] Countable Loop %" << CurLoop->getHeader()->getName() << "\n"); // The following transforms hoist stores/memsets into the loop pre-header. // Give up if the loop has instructions that may throw. SimpleLoopSafetyInfo SafetyInfo; SafetyInfo.computeLoopSafetyInfo(CurLoop); if (SafetyInfo.anyBlockMayThrow()) return false; bool MadeChange = false; // Scan all the blocks in the loop that are not in subloops. for (auto *BB : CurLoop->getBlocks()) { // Ignore blocks in subloops. if (LI->getLoopFor(BB) != CurLoop) continue; MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks); } return MadeChange; } static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) { const SCEVConstant *ConstStride = cast(StoreEv->getOperand(1)); return ConstStride->getAPInt(); } /// getMemSetPatternValue - If a strided store of the specified value is safe to /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should /// be passed in. Otherwise, return null. /// /// Note that we don't ever attempt to use memset_pattern8 or 4, because these /// just replicate their input array and then pass on to memset_pattern16. static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) { // FIXME: This could check for UndefValue because it can be merged into any // other valid pattern. // If the value isn't a constant, we can't promote it to being in a constant // array. We could theoretically do a store to an alloca or something, but // that doesn't seem worthwhile. Constant *C = dyn_cast(V); if (!C || isa(C)) return nullptr; // Only handle simple values that are a power of two bytes in size. uint64_t Size = DL->getTypeSizeInBits(V->getType()); if (Size == 0 || (Size & 7) || (Size & (Size - 1))) return nullptr; // Don't care enough about darwin/ppc to implement this. if (DL->isBigEndian()) return nullptr; // Convert to size in bytes. Size /= 8; // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see // if the top and bottom are the same (e.g. for vectors and large integers). if (Size > 16) return nullptr; // If the constant is exactly 16 bytes, just use it. if (Size == 16) return C; // Otherwise, we'll use an array of the constants. unsigned ArraySize = 16 / Size; ArrayType *AT = ArrayType::get(V->getType(), ArraySize); return ConstantArray::get(AT, std::vector(ArraySize, C)); } LoopIdiomRecognize::LegalStoreKind LoopIdiomRecognize::isLegalStore(StoreInst *SI) { // Don't touch volatile stores. if (SI->isVolatile()) return LegalStoreKind::None; // We only want simple or unordered-atomic stores. if (!SI->isUnordered()) return LegalStoreKind::None; // Avoid merging nontemporal stores. if (SI->getMetadata(LLVMContext::MD_nontemporal)) return LegalStoreKind::None; Value *StoredVal = SI->getValueOperand(); Value *StorePtr = SI->getPointerOperand(); // Don't convert stores of non-integral pointer types to memsets (which stores // integers). if (DL->isNonIntegralPointerType(StoredVal->getType()->getScalarType())) return LegalStoreKind::None; // Reject stores that are so large that they overflow an unsigned. // When storing out scalable vectors we bail out for now, since the code // below currently only works for constant strides. TypeSize SizeInBits = DL->getTypeSizeInBits(StoredVal->getType()); if (SizeInBits.isScalable() || (SizeInBits.getFixedValue() & 7) || (SizeInBits.getFixedValue() >> 32) != 0) return LegalStoreKind::None; // See if the pointer expression is an AddRec like {base,+,1} on the current // loop, which indicates a strided store. If we have something else, it's a // random store we can't handle. const SCEVAddRecExpr *StoreEv = dyn_cast(SE->getSCEV(StorePtr)); if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) return LegalStoreKind::None; // Check to see if we have a constant stride. if (!isa(StoreEv->getOperand(1))) return LegalStoreKind::None; // See if the store can be turned into a memset. // If the stored value is a byte-wise value (like i32 -1), then it may be // turned into a memset of i8 -1, assuming that all the consecutive bytes // are stored. A store of i32 0x01020304 can never be turned into a memset, // but it can be turned into memset_pattern if the target supports it. Value *SplatValue = isBytewiseValue(StoredVal, *DL); // Note: memset and memset_pattern on unordered-atomic is yet not supported bool UnorderedAtomic = SI->isUnordered() && !SI->isSimple(); // If we're allowed to form a memset, and the stored value would be // acceptable for memset, use it. if (!UnorderedAtomic && HasMemset && SplatValue && !DisableLIRP::Memset && // Verify that the stored value is loop invariant. If not, we can't // promote the memset. CurLoop->isLoopInvariant(SplatValue)) { // It looks like we can use SplatValue. return LegalStoreKind::Memset; } if (!UnorderedAtomic && HasMemsetPattern && !DisableLIRP::Memset && // Don't create memset_pattern16s with address spaces. StorePtr->getType()->getPointerAddressSpace() == 0 && getMemSetPatternValue(StoredVal, DL)) { // It looks like we can use PatternValue! return LegalStoreKind::MemsetPattern; } // Otherwise, see if the store can be turned into a memcpy. if (HasMemcpy && !DisableLIRP::Memcpy) { // Check to see if the stride matches the size of the store. If so, then we // know that every byte is touched in the loop. APInt Stride = getStoreStride(StoreEv); unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); if (StoreSize != Stride && StoreSize != -Stride) return LegalStoreKind::None; // The store must be feeding a non-volatile load. LoadInst *LI = dyn_cast(SI->getValueOperand()); // Only allow non-volatile loads if (!LI || LI->isVolatile()) return LegalStoreKind::None; // Only allow simple or unordered-atomic loads if (!LI->isUnordered()) return LegalStoreKind::None; // See if the pointer expression is an AddRec like {base,+,1} on the current // loop, which indicates a strided load. If we have something else, it's a // random load we can't handle. const SCEVAddRecExpr *LoadEv = dyn_cast(SE->getSCEV(LI->getPointerOperand())); if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine()) return LegalStoreKind::None; // The store and load must share the same stride. if (StoreEv->getOperand(1) != LoadEv->getOperand(1)) return LegalStoreKind::None; // Success. This store can be converted into a memcpy. UnorderedAtomic = UnorderedAtomic || LI->isAtomic(); return UnorderedAtomic ? LegalStoreKind::UnorderedAtomicMemcpy : LegalStoreKind::Memcpy; } // This store can't be transformed into a memset/memcpy. return LegalStoreKind::None; } void LoopIdiomRecognize::collectStores(BasicBlock *BB) { StoreRefsForMemset.clear(); StoreRefsForMemsetPattern.clear(); StoreRefsForMemcpy.clear(); for (Instruction &I : *BB) { StoreInst *SI = dyn_cast(&I); if (!SI) continue; // Make sure this is a strided store with a constant stride. switch (isLegalStore(SI)) { case LegalStoreKind::None: // Nothing to do break; case LegalStoreKind::Memset: { // Find the base pointer. Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); StoreRefsForMemset[Ptr].push_back(SI); } break; case LegalStoreKind::MemsetPattern: { // Find the base pointer. Value *Ptr = getUnderlyingObject(SI->getPointerOperand()); StoreRefsForMemsetPattern[Ptr].push_back(SI); } break; case LegalStoreKind::Memcpy: case LegalStoreKind::UnorderedAtomicMemcpy: StoreRefsForMemcpy.push_back(SI); break; default: assert(false && "unhandled return value"); break; } } } /// runOnLoopBlock - Process the specified block, which lives in a counted loop /// with the specified backedge count. This block is known to be in the current /// loop and not in any subloops. bool LoopIdiomRecognize::runOnLoopBlock( BasicBlock *BB, const SCEV *BECount, SmallVectorImpl &ExitBlocks) { // We can only promote stores in this block if they are unconditionally // executed in the loop. For a block to be unconditionally executed, it has // to dominate all the exit blocks of the loop. Verify this now. for (BasicBlock *ExitBlock : ExitBlocks) if (!DT->dominates(BB, ExitBlock)) return false; bool MadeChange = false; // Look for store instructions, which may be optimized to memset/memcpy. collectStores(BB); // Look for a single store or sets of stores with a common base, which can be // optimized into a memset (memset_pattern). The latter most commonly happens // with structs and handunrolled loops. for (auto &SL : StoreRefsForMemset) MadeChange |= processLoopStores(SL.second, BECount, ForMemset::Yes); for (auto &SL : StoreRefsForMemsetPattern) MadeChange |= processLoopStores(SL.second, BECount, ForMemset::No); // Optimize the store into a memcpy, if it feeds an similarly strided load. for (auto &SI : StoreRefsForMemcpy) MadeChange |= processLoopStoreOfLoopLoad(SI, BECount); MadeChange |= processLoopMemIntrinsic( BB, &LoopIdiomRecognize::processLoopMemCpy, BECount); MadeChange |= processLoopMemIntrinsic( BB, &LoopIdiomRecognize::processLoopMemSet, BECount); return MadeChange; } /// See if this store(s) can be promoted to a memset. bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl &SL, const SCEV *BECount, ForMemset For) { // Try to find consecutive stores that can be transformed into memsets. SetVector Heads, Tails; SmallDenseMap ConsecutiveChain; // Do a quadratic search on all of the given stores and find // all of the pairs of stores that follow each other. SmallVector IndexQueue; for (unsigned i = 0, e = SL.size(); i < e; ++i) { assert(SL[i]->isSimple() && "Expected only non-volatile stores."); Value *FirstStoredVal = SL[i]->getValueOperand(); Value *FirstStorePtr = SL[i]->getPointerOperand(); const SCEVAddRecExpr *FirstStoreEv = cast(SE->getSCEV(FirstStorePtr)); APInt FirstStride = getStoreStride(FirstStoreEv); unsigned FirstStoreSize = DL->getTypeStoreSize(SL[i]->getValueOperand()->getType()); // See if we can optimize just this store in isolation. if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) { Heads.insert(SL[i]); continue; } Value *FirstSplatValue = nullptr; Constant *FirstPatternValue = nullptr; if (For == ForMemset::Yes) FirstSplatValue = isBytewiseValue(FirstStoredVal, *DL); else FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL); assert((FirstSplatValue || FirstPatternValue) && "Expected either splat value or pattern value."); IndexQueue.clear(); // If a store has multiple consecutive store candidates, search Stores // array according to the sequence: from i+1 to e, then from i-1 to 0. // This is because usually pairing with immediate succeeding or preceding // candidate create the best chance to find memset opportunity. unsigned j = 0; for (j = i + 1; j < e; ++j) IndexQueue.push_back(j); for (j = i; j > 0; --j) IndexQueue.push_back(j - 1); for (auto &k : IndexQueue) { assert(SL[k]->isSimple() && "Expected only non-volatile stores."); Value *SecondStorePtr = SL[k]->getPointerOperand(); const SCEVAddRecExpr *SecondStoreEv = cast(SE->getSCEV(SecondStorePtr)); APInt SecondStride = getStoreStride(SecondStoreEv); if (FirstStride != SecondStride) continue; Value *SecondStoredVal = SL[k]->getValueOperand(); Value *SecondSplatValue = nullptr; Constant *SecondPatternValue = nullptr; if (For == ForMemset::Yes) SecondSplatValue = isBytewiseValue(SecondStoredVal, *DL); else SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL); assert((SecondSplatValue || SecondPatternValue) && "Expected either splat value or pattern value."); if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) { if (For == ForMemset::Yes) { if (isa(FirstSplatValue)) FirstSplatValue = SecondSplatValue; if (FirstSplatValue != SecondSplatValue) continue; } else { if (isa(FirstPatternValue)) FirstPatternValue = SecondPatternValue; if (FirstPatternValue != SecondPatternValue) continue; } Tails.insert(SL[k]); Heads.insert(SL[i]); ConsecutiveChain[SL[i]] = SL[k]; break; } } } // We may run into multiple chains that merge into a single chain. We mark the // stores that we transformed so that we don't visit the same store twice. SmallPtrSet TransformedStores; bool Changed = false; // For stores that start but don't end a link in the chain: for (StoreInst *I : Heads) { if (Tails.count(I)) continue; // We found a store instr that starts a chain. Now follow the chain and try // to transform it. SmallPtrSet AdjacentStores; StoreInst *HeadStore = I; unsigned StoreSize = 0; // Collect the chain into a list. while (Tails.count(I) || Heads.count(I)) { if (TransformedStores.count(I)) break; AdjacentStores.insert(I); StoreSize += DL->getTypeStoreSize(I->getValueOperand()->getType()); // Move to the next value in the chain. I = ConsecutiveChain[I]; } Value *StoredVal = HeadStore->getValueOperand(); Value *StorePtr = HeadStore->getPointerOperand(); const SCEVAddRecExpr *StoreEv = cast(SE->getSCEV(StorePtr)); APInt Stride = getStoreStride(StoreEv); // Check to see if the stride matches the size of the stores. If so, then // we know that every byte is touched in the loop. if (StoreSize != Stride && StoreSize != -Stride) continue; bool IsNegStride = StoreSize == -Stride; Type *IntIdxTy = DL->getIndexType(StorePtr->getType()); const SCEV *StoreSizeSCEV = SE->getConstant(IntIdxTy, StoreSize); if (processLoopStridedStore(StorePtr, StoreSizeSCEV, MaybeAlign(HeadStore->getAlign()), StoredVal, HeadStore, AdjacentStores, StoreEv, BECount, IsNegStride)) { TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end()); Changed = true; } } return Changed; } /// processLoopMemIntrinsic - Template function for calling different processor /// functions based on mem intrinsic type. template bool LoopIdiomRecognize::processLoopMemIntrinsic( BasicBlock *BB, bool (LoopIdiomRecognize::*Processor)(MemInst *, const SCEV *), const SCEV *BECount) { bool MadeChange = false; for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) { Instruction *Inst = &*I++; // Look for memory instructions, which may be optimized to a larger one. if (MemInst *MI = dyn_cast(Inst)) { WeakTrackingVH InstPtr(&*I); if (!(this->*Processor)(MI, BECount)) continue; MadeChange = true; // If processing the instruction invalidated our iterator, start over from // the top of the block. if (!InstPtr) I = BB->begin(); } } return MadeChange; } /// processLoopMemCpy - See if this memcpy can be promoted to a large memcpy bool LoopIdiomRecognize::processLoopMemCpy(MemCpyInst *MCI, const SCEV *BECount) { // We can only handle non-volatile memcpys with a constant size. if (MCI->isVolatile() || !isa(MCI->getLength())) return false; // If we're not allowed to hack on memcpy, we fail. if ((!HasMemcpy && !isa(MCI)) || DisableLIRP::Memcpy) return false; Value *Dest = MCI->getDest(); Value *Source = MCI->getSource(); if (!Dest || !Source) return false; // See if the load and store pointer expressions are AddRec like {base,+,1} on // the current loop, which indicates a strided load and store. If we have // something else, it's a random load or store we can't handle. const SCEVAddRecExpr *StoreEv = dyn_cast(SE->getSCEV(Dest)); if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) return false; const SCEVAddRecExpr *LoadEv = dyn_cast(SE->getSCEV(Source)); if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine()) return false; // Reject memcpys that are so large that they overflow an unsigned. uint64_t SizeInBytes = cast(MCI->getLength())->getZExtValue(); if ((SizeInBytes >> 32) != 0) return false; // Check if the stride matches the size of the memcpy. If so, then we know // that every byte is touched in the loop. const SCEVConstant *ConstStoreStride = dyn_cast(StoreEv->getOperand(1)); const SCEVConstant *ConstLoadStride = dyn_cast(LoadEv->getOperand(1)); if (!ConstStoreStride || !ConstLoadStride) return false; APInt StoreStrideValue = ConstStoreStride->getAPInt(); APInt LoadStrideValue = ConstLoadStride->getAPInt(); // Huge stride value - give up if (StoreStrideValue.getBitWidth() > 64 || LoadStrideValue.getBitWidth() > 64) return false; if (SizeInBytes != StoreStrideValue && SizeInBytes != -StoreStrideValue) { ORE.emit([&]() { return OptimizationRemarkMissed(DEBUG_TYPE, "SizeStrideUnequal", MCI) << ore::NV("Inst", "memcpy") << " in " << ore::NV("Function", MCI->getFunction()) << " function will not be hoisted: " << ore::NV("Reason", "memcpy size is not equal to stride"); }); return false; } int64_t StoreStrideInt = StoreStrideValue.getSExtValue(); int64_t LoadStrideInt = LoadStrideValue.getSExtValue(); // Check if the load stride matches the store stride. if (StoreStrideInt != LoadStrideInt) return false; return processLoopStoreOfLoopLoad( Dest, Source, SE->getConstant(Dest->getType(), SizeInBytes), MCI->getDestAlign(), MCI->getSourceAlign(), MCI, MCI, StoreEv, LoadEv, BECount); } /// processLoopMemSet - See if this memset can be promoted to a large memset. bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI, const SCEV *BECount) { // We can only handle non-volatile memsets. if (MSI->isVolatile()) return false; // If we're not allowed to hack on memset, we fail. if (!HasMemset || DisableLIRP::Memset) return false; Value *Pointer = MSI->getDest(); // See if the pointer expression is an AddRec like {base,+,1} on the current // loop, which indicates a strided store. If we have something else, it's a // random store we can't handle. const SCEVAddRecExpr *Ev = dyn_cast(SE->getSCEV(Pointer)); if (!Ev || Ev->getLoop() != CurLoop) return false; if (!Ev->isAffine()) { LLVM_DEBUG(dbgs() << " Pointer is not affine, abort\n"); return false; } const SCEV *PointerStrideSCEV = Ev->getOperand(1); const SCEV *MemsetSizeSCEV = SE->getSCEV(MSI->getLength()); if (!PointerStrideSCEV || !MemsetSizeSCEV) return false; bool IsNegStride = false; const bool IsConstantSize = isa(MSI->getLength()); if (IsConstantSize) { // Memset size is constant. // Check if the pointer stride matches the memset size. If so, then // we know that every byte is touched in the loop. LLVM_DEBUG(dbgs() << " memset size is constant\n"); uint64_t SizeInBytes = cast(MSI->getLength())->getZExtValue(); const SCEVConstant *ConstStride = dyn_cast(Ev->getOperand(1)); if (!ConstStride) return false; APInt Stride = ConstStride->getAPInt(); if (SizeInBytes != Stride && SizeInBytes != -Stride) return false; IsNegStride = SizeInBytes == -Stride; } else { // Memset size is non-constant. // Check if the pointer stride matches the memset size. // To be conservative, the pass would not promote pointers that aren't in // address space zero. Also, the pass only handles memset length and stride // that are invariant for the top level loop. LLVM_DEBUG(dbgs() << " memset size is non-constant\n"); if (Pointer->getType()->getPointerAddressSpace() != 0) { LLVM_DEBUG(dbgs() << " pointer is not in address space zero, " << "abort\n"); return false; } if (!SE->isLoopInvariant(MemsetSizeSCEV, CurLoop)) { LLVM_DEBUG(dbgs() << " memset size is not a loop-invariant, " << "abort\n"); return false; } // Compare positive direction PointerStrideSCEV with MemsetSizeSCEV IsNegStride = PointerStrideSCEV->isNonConstantNegative(); const SCEV *PositiveStrideSCEV = IsNegStride ? SE->getNegativeSCEV(PointerStrideSCEV) : PointerStrideSCEV; LLVM_DEBUG(dbgs() << " MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n" << " PositiveStrideSCEV: " << *PositiveStrideSCEV << "\n"); if (PositiveStrideSCEV != MemsetSizeSCEV) { // If an expression is covered by the loop guard, compare again and // proceed with optimization if equal. const SCEV *FoldedPositiveStride = SE->applyLoopGuards(PositiveStrideSCEV, CurLoop); const SCEV *FoldedMemsetSize = SE->applyLoopGuards(MemsetSizeSCEV, CurLoop); LLVM_DEBUG(dbgs() << " Try to fold SCEV based on loop guard\n" << " FoldedMemsetSize: " << *FoldedMemsetSize << "\n" << " FoldedPositiveStride: " << *FoldedPositiveStride << "\n"); if (FoldedPositiveStride != FoldedMemsetSize) { LLVM_DEBUG(dbgs() << " SCEV don't match, abort\n"); return false; } } } // Verify that the memset value is loop invariant. If not, we can't promote // the memset. Value *SplatValue = MSI->getValue(); if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue)) return false; SmallPtrSet MSIs; MSIs.insert(MSI); return processLoopStridedStore(Pointer, SE->getSCEV(MSI->getLength()), MSI->getDestAlign(), SplatValue, MSI, MSIs, Ev, BECount, IsNegStride, /*IsLoopMemset=*/true); } /// mayLoopAccessLocation - Return true if the specified loop might access the /// specified pointer location, which is a loop-strided access. The 'Access' /// argument specifies what the verboten forms of access are (read or write). static bool mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L, const SCEV *BECount, const SCEV *StoreSizeSCEV, AliasAnalysis &AA, SmallPtrSetImpl &IgnoredInsts) { // Get the location that may be stored across the loop. Since the access is // strided positively through memory, we say that the modified location starts // at the pointer and has infinite size. LocationSize AccessSize = LocationSize::afterPointer(); // If the loop iterates a fixed number of times, we can refine the access size // to be exactly the size of the memset, which is (BECount+1)*StoreSize const SCEVConstant *BECst = dyn_cast(BECount); const SCEVConstant *ConstSize = dyn_cast(StoreSizeSCEV); if (BECst && ConstSize) AccessSize = LocationSize::precise((BECst->getValue()->getZExtValue() + 1) * ConstSize->getValue()->getZExtValue()); // TODO: For this to be really effective, we have to dive into the pointer // operand in the store. Store to &A[i] of 100 will always return may alias // with store of &A[100], we need to StoreLoc to be "A" with size of 100, // which will then no-alias a store to &A[100]. MemoryLocation StoreLoc(Ptr, AccessSize); for (BasicBlock *B : L->blocks()) for (Instruction &I : *B) if (!IgnoredInsts.contains(&I) && isModOrRefSet(AA.getModRefInfo(&I, StoreLoc) & Access)) return true; return false; } // If we have a negative stride, Start refers to the end of the memory location // we're trying to memset. Therefore, we need to recompute the base pointer, // which is just Start - BECount*Size. static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount, Type *IntPtr, const SCEV *StoreSizeSCEV, ScalarEvolution *SE) { const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr); if (!StoreSizeSCEV->isOne()) { // index = back edge count * store size Index = SE->getMulExpr(Index, SE->getTruncateOrZeroExtend(StoreSizeSCEV, IntPtr), SCEV::FlagNUW); } // base pointer = start - index * store size return SE->getMinusSCEV(Start, Index); } /// Compute the number of bytes as a SCEV from the backedge taken count. /// /// This also maps the SCEV into the provided type and tries to handle the /// computation in a way that will fold cleanly. static const SCEV *getNumBytes(const SCEV *BECount, Type *IntPtr, const SCEV *StoreSizeSCEV, Loop *CurLoop, const DataLayout *DL, ScalarEvolution *SE) { const SCEV *TripCountSCEV = SE->getTripCountFromExitCount(BECount, IntPtr, CurLoop); return SE->getMulExpr(TripCountSCEV, SE->getTruncateOrZeroExtend(StoreSizeSCEV, IntPtr), SCEV::FlagNUW); } /// processLoopStridedStore - We see a strided store of some value. If we can /// transform this into a memset or memset_pattern in the loop preheader, do so. bool LoopIdiomRecognize::processLoopStridedStore( Value *DestPtr, const SCEV *StoreSizeSCEV, MaybeAlign StoreAlignment, Value *StoredVal, Instruction *TheStore, SmallPtrSetImpl &Stores, const SCEVAddRecExpr *Ev, const SCEV *BECount, bool IsNegStride, bool IsLoopMemset) { Module *M = TheStore->getModule(); Value *SplatValue = isBytewiseValue(StoredVal, *DL); Constant *PatternValue = nullptr; if (!SplatValue) PatternValue = getMemSetPatternValue(StoredVal, DL); assert((SplatValue || PatternValue) && "Expected either splat value or pattern value."); // The trip count of the loop and the base pointer of the addrec SCEV is // guaranteed to be loop invariant, which means that it should dominate the // header. This allows us to insert code for it in the preheader. unsigned DestAS = DestPtr->getType()->getPointerAddressSpace(); BasicBlock *Preheader = CurLoop->getLoopPreheader(); IRBuilder<> Builder(Preheader->getTerminator()); SCEVExpander Expander(*SE, *DL, "loop-idiom"); SCEVExpanderCleaner ExpCleaner(Expander); Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS); Type *IntIdxTy = DL->getIndexType(DestPtr->getType()); bool Changed = false; const SCEV *Start = Ev->getStart(); // Handle negative strided loops. if (IsNegStride) Start = getStartForNegStride(Start, BECount, IntIdxTy, StoreSizeSCEV, SE); // TODO: ideally we should still be able to generate memset if SCEV expander // is taught to generate the dependencies at the latest point. if (!Expander.isSafeToExpand(Start)) return Changed; // Okay, we have a strided store "p[i]" of a splattable value. We can turn // this into a memset in the loop preheader now if we want. However, this // would be unsafe to do if there is anything else in the loop that may read // or write to the aliased location. Check for any overlap by generating the // base pointer and checking the region. Value *BasePtr = Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator()); // From here on out, conservatively report to the pass manager that we've // changed the IR, even if we later clean up these added instructions. There // may be structural differences e.g. in the order of use lists not accounted // for in just a textual dump of the IR. This is written as a variable, even // though statically all the places this dominates could be replaced with // 'true', with the hope that anyone trying to be clever / "more precise" with // the return value will read this comment, and leave them alone. Changed = true; if (mayLoopAccessLocation(BasePtr, ModRefInfo::ModRef, CurLoop, BECount, StoreSizeSCEV, *AA, Stores)) return Changed; if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset)) return Changed; // Okay, everything looks good, insert the memset. const SCEV *NumBytesS = getNumBytes(BECount, IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE); // TODO: ideally we should still be able to generate memset if SCEV expander // is taught to generate the dependencies at the latest point. if (!Expander.isSafeToExpand(NumBytesS)) return Changed; Value *NumBytes = Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); if (!SplatValue && !isLibFuncEmittable(M, TLI, LibFunc_memset_pattern16)) return Changed; AAMDNodes AATags = TheStore->getAAMetadata(); for (Instruction *Store : Stores) AATags = AATags.merge(Store->getAAMetadata()); if (auto CI = dyn_cast(NumBytes)) AATags = AATags.extendTo(CI->getZExtValue()); else AATags = AATags.extendTo(-1); CallInst *NewCall; if (SplatValue) { NewCall = Builder.CreateMemSet( BasePtr, SplatValue, NumBytes, MaybeAlign(StoreAlignment), /*isVolatile=*/false, AATags.TBAA, AATags.Scope, AATags.NoAlias); } else { assert (isLibFuncEmittable(M, TLI, LibFunc_memset_pattern16)); // Everything is emitted in default address space Type *Int8PtrTy = DestInt8PtrTy; StringRef FuncName = "memset_pattern16"; FunctionCallee MSP = getOrInsertLibFunc(M, *TLI, LibFunc_memset_pattern16, Builder.getVoidTy(), Int8PtrTy, Int8PtrTy, IntIdxTy); inferNonMandatoryLibFuncAttrs(M, FuncName, *TLI); // Otherwise we should form a memset_pattern16. PatternValue is known to be // an constant array of 16-bytes. Plop the value into a mergable global. GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true, GlobalValue::PrivateLinkage, PatternValue, ".memset_pattern"); GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these. GV->setAlignment(Align(16)); Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy); NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes}); // Set the TBAA info if present. if (AATags.TBAA) NewCall->setMetadata(LLVMContext::MD_tbaa, AATags.TBAA); if (AATags.Scope) NewCall->setMetadata(LLVMContext::MD_alias_scope, AATags.Scope); if (AATags.NoAlias) NewCall->setMetadata(LLVMContext::MD_noalias, AATags.NoAlias); } NewCall->setDebugLoc(TheStore->getDebugLoc()); if (MSSAU) { MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); MSSAU->insertDef(cast(NewMemAcc), true); } LLVM_DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n" << " from store to: " << *Ev << " at: " << *TheStore << "\n"); ORE.emit([&]() { OptimizationRemark R(DEBUG_TYPE, "ProcessLoopStridedStore", NewCall->getDebugLoc(), Preheader); R << "Transformed loop-strided store in " << ore::NV("Function", TheStore->getFunction()) << " function into a call to " << ore::NV("NewFunction", NewCall->getCalledFunction()) << "() intrinsic"; if (!Stores.empty()) R << ore::setExtraArgs(); for (auto *I : Stores) { R << ore::NV("FromBlock", I->getParent()->getName()) << ore::NV("ToBlock", Preheader->getName()); } return R; }); // Okay, the memset has been formed. Zap the original store and anything that // feeds into it. for (auto *I : Stores) { if (MSSAU) MSSAU->removeMemoryAccess(I, true); deleteDeadInstruction(I); } if (MSSAU && VerifyMemorySSA) MSSAU->getMemorySSA()->verifyMemorySSA(); ++NumMemSet; ExpCleaner.markResultUsed(); return true; } /// If the stored value is a strided load in the same loop with the same stride /// this may be transformable into a memcpy. This kicks in for stuff like /// for (i) A[i] = B[i]; bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount) { assert(SI->isUnordered() && "Expected only non-volatile non-ordered stores."); Value *StorePtr = SI->getPointerOperand(); const SCEVAddRecExpr *StoreEv = cast(SE->getSCEV(StorePtr)); unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType()); // The store must be feeding a non-volatile load. LoadInst *LI = cast(SI->getValueOperand()); assert(LI->isUnordered() && "Expected only non-volatile non-ordered loads."); // See if the pointer expression is an AddRec like {base,+,1} on the current // loop, which indicates a strided load. If we have something else, it's a // random load we can't handle. Value *LoadPtr = LI->getPointerOperand(); const SCEVAddRecExpr *LoadEv = cast(SE->getSCEV(LoadPtr)); const SCEV *StoreSizeSCEV = SE->getConstant(StorePtr->getType(), StoreSize); return processLoopStoreOfLoopLoad(StorePtr, LoadPtr, StoreSizeSCEV, SI->getAlign(), LI->getAlign(), SI, LI, StoreEv, LoadEv, BECount); } namespace { class MemmoveVerifier { public: explicit MemmoveVerifier(const Value &LoadBasePtr, const Value &StoreBasePtr, const DataLayout &DL) : DL(DL), BP1(llvm::GetPointerBaseWithConstantOffset( LoadBasePtr.stripPointerCasts(), LoadOff, DL)), BP2(llvm::GetPointerBaseWithConstantOffset( StoreBasePtr.stripPointerCasts(), StoreOff, DL)), IsSameObject(BP1 == BP2) {} bool loadAndStoreMayFormMemmove(unsigned StoreSize, bool IsNegStride, const Instruction &TheLoad, bool IsMemCpy) const { if (IsMemCpy) { // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr // for negative stride. if ((!IsNegStride && LoadOff <= StoreOff) || (IsNegStride && LoadOff >= StoreOff)) return false; } else { // Ensure that LoadBasePtr is after StoreBasePtr or before StoreBasePtr // for negative stride. LoadBasePtr shouldn't overlap with StoreBasePtr. int64_t LoadSize = DL.getTypeSizeInBits(TheLoad.getType()).getFixedValue() / 8; if (BP1 != BP2 || LoadSize != int64_t(StoreSize)) return false; if ((!IsNegStride && LoadOff < StoreOff + int64_t(StoreSize)) || (IsNegStride && LoadOff + LoadSize > StoreOff)) return false; } return true; } private: const DataLayout &DL; int64_t LoadOff = 0; int64_t StoreOff = 0; const Value *BP1; const Value *BP2; public: const bool IsSameObject; }; } // namespace bool LoopIdiomRecognize::processLoopStoreOfLoopLoad( Value *DestPtr, Value *SourcePtr, const SCEV *StoreSizeSCEV, MaybeAlign StoreAlign, MaybeAlign LoadAlign, Instruction *TheStore, Instruction *TheLoad, const SCEVAddRecExpr *StoreEv, const SCEVAddRecExpr *LoadEv, const SCEV *BECount) { // FIXME: until llvm.memcpy.inline supports dynamic sizes, we need to // conservatively bail here, since otherwise we may have to transform // llvm.memcpy.inline into llvm.memcpy which is illegal. if (isa(TheStore)) return false; // The trip count of the loop and the base pointer of the addrec SCEV is // guaranteed to be loop invariant, which means that it should dominate the // header. This allows us to insert code for it in the preheader. BasicBlock *Preheader = CurLoop->getLoopPreheader(); IRBuilder<> Builder(Preheader->getTerminator()); SCEVExpander Expander(*SE, *DL, "loop-idiom"); SCEVExpanderCleaner ExpCleaner(Expander); bool Changed = false; const SCEV *StrStart = StoreEv->getStart(); unsigned StrAS = DestPtr->getType()->getPointerAddressSpace(); Type *IntIdxTy = Builder.getIntNTy(DL->getIndexSizeInBits(StrAS)); APInt Stride = getStoreStride(StoreEv); const SCEVConstant *ConstStoreSize = dyn_cast(StoreSizeSCEV); // TODO: Deal with non-constant size; Currently expect constant store size assert(ConstStoreSize && "store size is expected to be a constant"); int64_t StoreSize = ConstStoreSize->getValue()->getZExtValue(); bool IsNegStride = StoreSize == -Stride; // Handle negative strided loops. if (IsNegStride) StrStart = getStartForNegStride(StrStart, BECount, IntIdxTy, StoreSizeSCEV, SE); // Okay, we have a strided store "p[i]" of a loaded value. We can turn // this into a memcpy in the loop preheader now if we want. However, this // would be unsafe to do if there is anything else in the loop that may read // or write the memory region we're storing to. This includes the load that // feeds the stores. Check for an alias by generating the base address and // checking everything. Value *StoreBasePtr = Expander.expandCodeFor( StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator()); // From here on out, conservatively report to the pass manager that we've // changed the IR, even if we later clean up these added instructions. There // may be structural differences e.g. in the order of use lists not accounted // for in just a textual dump of the IR. This is written as a variable, even // though statically all the places this dominates could be replaced with // 'true', with the hope that anyone trying to be clever / "more precise" with // the return value will read this comment, and leave them alone. Changed = true; SmallPtrSet IgnoredInsts; IgnoredInsts.insert(TheStore); bool IsMemCpy = isa(TheStore); const StringRef InstRemark = IsMemCpy ? "memcpy" : "load and store"; bool LoopAccessStore = mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount, StoreSizeSCEV, *AA, IgnoredInsts); if (LoopAccessStore) { // For memmove case it's not enough to guarantee that loop doesn't access // TheStore and TheLoad. Additionally we need to make sure that TheStore is // the only user of TheLoad. if (!TheLoad->hasOneUse()) return Changed; IgnoredInsts.insert(TheLoad); if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount, StoreSizeSCEV, *AA, IgnoredInsts)) { ORE.emit([&]() { return OptimizationRemarkMissed(DEBUG_TYPE, "LoopMayAccessStore", TheStore) << ore::NV("Inst", InstRemark) << " in " << ore::NV("Function", TheStore->getFunction()) << " function will not be hoisted: " << ore::NV("Reason", "The loop may access store location"); }); return Changed; } IgnoredInsts.erase(TheLoad); } const SCEV *LdStart = LoadEv->getStart(); unsigned LdAS = SourcePtr->getType()->getPointerAddressSpace(); // Handle negative strided loops. if (IsNegStride) LdStart = getStartForNegStride(LdStart, BECount, IntIdxTy, StoreSizeSCEV, SE); // For a memcpy, we have to make sure that the input array is not being // mutated by the loop. Value *LoadBasePtr = Expander.expandCodeFor( LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator()); // If the store is a memcpy instruction, we must check if it will write to // the load memory locations. So remove it from the ignored stores. MemmoveVerifier Verifier(*LoadBasePtr, *StoreBasePtr, *DL); if (IsMemCpy && !Verifier.IsSameObject) IgnoredInsts.erase(TheStore); if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount, StoreSizeSCEV, *AA, IgnoredInsts)) { ORE.emit([&]() { return OptimizationRemarkMissed(DEBUG_TYPE, "LoopMayAccessLoad", TheLoad) << ore::NV("Inst", InstRemark) << " in " << ore::NV("Function", TheStore->getFunction()) << " function will not be hoisted: " << ore::NV("Reason", "The loop may access load location"); }); return Changed; } bool UseMemMove = IsMemCpy ? Verifier.IsSameObject : LoopAccessStore; if (UseMemMove) if (!Verifier.loadAndStoreMayFormMemmove(StoreSize, IsNegStride, *TheLoad, IsMemCpy)) return Changed; if (avoidLIRForMultiBlockLoop()) return Changed; // Okay, everything is safe, we can transform this! const SCEV *NumBytesS = getNumBytes(BECount, IntIdxTy, StoreSizeSCEV, CurLoop, DL, SE); Value *NumBytes = Expander.expandCodeFor(NumBytesS, IntIdxTy, Preheader->getTerminator()); AAMDNodes AATags = TheLoad->getAAMetadata(); AAMDNodes StoreAATags = TheStore->getAAMetadata(); AATags = AATags.merge(StoreAATags); if (auto CI = dyn_cast(NumBytes)) AATags = AATags.extendTo(CI->getZExtValue()); else AATags = AATags.extendTo(-1); CallInst *NewCall = nullptr; // Check whether to generate an unordered atomic memcpy: // If the load or store are atomic, then they must necessarily be unordered // by previous checks. if (!TheStore->isAtomic() && !TheLoad->isAtomic()) { if (UseMemMove) NewCall = Builder.CreateMemMove( StoreBasePtr, StoreAlign, LoadBasePtr, LoadAlign, NumBytes, /*isVolatile=*/false, AATags.TBAA, AATags.Scope, AATags.NoAlias); else NewCall = Builder.CreateMemCpy(StoreBasePtr, StoreAlign, LoadBasePtr, LoadAlign, NumBytes, /*isVolatile=*/false, AATags.TBAA, AATags.TBAAStruct, AATags.Scope, AATags.NoAlias); } else { // For now don't support unordered atomic memmove. if (UseMemMove) return Changed; // We cannot allow unaligned ops for unordered load/store, so reject // anything where the alignment isn't at least the element size. assert((StoreAlign && LoadAlign) && "Expect unordered load/store to have align."); if (*StoreAlign < StoreSize || *LoadAlign < StoreSize) return Changed; // If the element.atomic memcpy is not lowered into explicit // loads/stores later, then it will be lowered into an element-size // specific lib call. If the lib call doesn't exist for our store size, then // we shouldn't generate the memcpy. if (StoreSize > TTI->getAtomicMemIntrinsicMaxElementSize()) return Changed; // Create the call. // Note that unordered atomic loads/stores are *required* by the spec to // have an alignment but non-atomic loads/stores may not. NewCall = Builder.CreateElementUnorderedAtomicMemCpy( StoreBasePtr, *StoreAlign, LoadBasePtr, *LoadAlign, NumBytes, StoreSize, AATags.TBAA, AATags.TBAAStruct, AATags.Scope, AATags.NoAlias); } NewCall->setDebugLoc(TheStore->getDebugLoc()); if (MSSAU) { MemoryAccess *NewMemAcc = MSSAU->createMemoryAccessInBB( NewCall, nullptr, NewCall->getParent(), MemorySSA::BeforeTerminator); MSSAU->insertDef(cast(NewMemAcc), true); } LLVM_DEBUG(dbgs() << " Formed new call: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *TheLoad << "\n" << " from store ptr=" << *StoreEv << " at: " << *TheStore << "\n"); ORE.emit([&]() { return OptimizationRemark(DEBUG_TYPE, "ProcessLoopStoreOfLoopLoad", NewCall->getDebugLoc(), Preheader) << "Formed a call to " << ore::NV("NewFunction", NewCall->getCalledFunction()) << "() intrinsic from " << ore::NV("Inst", InstRemark) << " instruction in " << ore::NV("Function", TheStore->getFunction()) << " function" << ore::setExtraArgs() << ore::NV("FromBlock", TheStore->getParent()->getName()) << ore::NV("ToBlock", Preheader->getName()); }); // Okay, a new call to memcpy/memmove has been formed. Zap the original store // and anything that feeds into it. if (MSSAU) MSSAU->removeMemoryAccess(TheStore, true); deleteDeadInstruction(TheStore); if (MSSAU && VerifyMemorySSA) MSSAU->getMemorySSA()->verifyMemorySSA(); if (UseMemMove) ++NumMemMove; else ++NumMemCpy; ExpCleaner.markResultUsed(); return true; } // When compiling for codesize we avoid idiom recognition for a multi-block loop // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop. // bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset, bool IsLoopMemset) { if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) { if (CurLoop->isOutermost() && (!IsMemset || !IsLoopMemset)) { LLVM_DEBUG(dbgs() << " " << CurLoop->getHeader()->getParent()->getName() << " : LIR " << (IsMemset ? "Memset" : "Memcpy") << " avoided: multi-block top-level loop\n"); return true; } } return false; } bool LoopIdiomRecognize::runOnNoncountableLoop() { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Scanning: F[" << CurLoop->getHeader()->getParent()->getName() << "] Noncountable Loop %" << CurLoop->getHeader()->getName() << "\n"); return recognizePopcount() || recognizeAndInsertFFS() || recognizeShiftUntilBitTest() || recognizeShiftUntilZero(); } /// Check if the given conditional branch is based on the comparison between /// a variable and zero, and if the variable is non-zero or zero (JmpOnZero is /// true), the control yields to the loop entry. If the branch matches the /// behavior, the variable involved in the comparison is returned. This function /// will be called to see if the precondition and postcondition of the loop are /// in desirable form. static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry, bool JmpOnZero = false) { if (!BI || !BI->isConditional()) return nullptr; ICmpInst *Cond = dyn_cast(BI->getCondition()); if (!Cond) return nullptr; ConstantInt *CmpZero = dyn_cast(Cond->getOperand(1)); if (!CmpZero || !CmpZero->isZero()) return nullptr; BasicBlock *TrueSucc = BI->getSuccessor(0); BasicBlock *FalseSucc = BI->getSuccessor(1); if (JmpOnZero) std::swap(TrueSucc, FalseSucc); ICmpInst::Predicate Pred = Cond->getPredicate(); if ((Pred == ICmpInst::ICMP_NE && TrueSucc == LoopEntry) || (Pred == ICmpInst::ICMP_EQ && FalseSucc == LoopEntry)) return Cond->getOperand(0); return nullptr; } // Check if the recurrence variable `VarX` is in the right form to create // the idiom. Returns the value coerced to a PHINode if so. static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX, BasicBlock *LoopEntry) { auto *PhiX = dyn_cast(VarX); if (PhiX && PhiX->getParent() == LoopEntry && (PhiX->getOperand(0) == DefX || PhiX->getOperand(1) == DefX)) return PhiX; return nullptr; } /// Return true iff the idiom is detected in the loop. /// /// Additionally: /// 1) \p CntInst is set to the instruction counting the population bit. /// 2) \p CntPhi is set to the corresponding phi node. /// 3) \p Var is set to the value whose population bits are being counted. /// /// The core idiom we are trying to detect is: /// \code /// if (x0 != 0) /// goto loop-exit // the precondition of the loop /// cnt0 = init-val; /// do { /// x1 = phi (x0, x2); /// cnt1 = phi(cnt0, cnt2); /// /// cnt2 = cnt1 + 1; /// ... /// x2 = x1 & (x1 - 1); /// ... /// } while(x != 0); /// /// loop-exit: /// \endcode static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB, Instruction *&CntInst, PHINode *&CntPhi, Value *&Var) { // step 1: Check to see if the look-back branch match this pattern: // "if (a!=0) goto loop-entry". BasicBlock *LoopEntry; Instruction *DefX2, *CountInst; Value *VarX1, *VarX0; PHINode *PhiX, *CountPhi; DefX2 = CountInst = nullptr; VarX1 = VarX0 = nullptr; PhiX = CountPhi = nullptr; LoopEntry = *(CurLoop->block_begin()); // step 1: Check if the loop-back branch is in desirable form. { if (Value *T = matchCondition( dyn_cast(LoopEntry->getTerminator()), LoopEntry)) DefX2 = dyn_cast(T); else return false; } // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)" { if (!DefX2 || DefX2->getOpcode() != Instruction::And) return false; BinaryOperator *SubOneOp; if ((SubOneOp = dyn_cast(DefX2->getOperand(0)))) VarX1 = DefX2->getOperand(1); else { VarX1 = DefX2->getOperand(0); SubOneOp = dyn_cast(DefX2->getOperand(1)); } if (!SubOneOp || SubOneOp->getOperand(0) != VarX1) return false; ConstantInt *Dec = dyn_cast(SubOneOp->getOperand(1)); if (!Dec || !((SubOneOp->getOpcode() == Instruction::Sub && Dec->isOne()) || (SubOneOp->getOpcode() == Instruction::Add && Dec->isMinusOne()))) { return false; } } // step 3: Check the recurrence of variable X PhiX = getRecurrenceVar(VarX1, DefX2, LoopEntry); if (!PhiX) return false; // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1 { CountInst = nullptr; for (Instruction &Inst : llvm::make_range( LoopEntry->getFirstNonPHI()->getIterator(), LoopEntry->end())) { if (Inst.getOpcode() != Instruction::Add) continue; ConstantInt *Inc = dyn_cast(Inst.getOperand(1)); if (!Inc || !Inc->isOne()) continue; PHINode *Phi = getRecurrenceVar(Inst.getOperand(0), &Inst, LoopEntry); if (!Phi) continue; // Check if the result of the instruction is live of the loop. bool LiveOutLoop = false; for (User *U : Inst.users()) { if ((cast(U))->getParent() != LoopEntry) { LiveOutLoop = true; break; } } if (LiveOutLoop) { CountInst = &Inst; CountPhi = Phi; break; } } if (!CountInst) return false; } // step 5: check if the precondition is in this form: // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;" { auto *PreCondBr = dyn_cast(PreCondBB->getTerminator()); Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader()); if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1)) return false; CntInst = CountInst; CntPhi = CountPhi; Var = T; } return true; } /// Return true if the idiom is detected in the loop. /// /// Additionally: /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ) /// or nullptr if there is no such. /// 2) \p CntPhi is set to the corresponding phi node /// or nullptr if there is no such. /// 3) \p Var is set to the value whose CTLZ could be used. /// 4) \p DefX is set to the instruction calculating Loop exit condition. /// /// The core idiom we are trying to detect is: /// \code /// if (x0 == 0) /// goto loop-exit // the precondition of the loop /// cnt0 = init-val; /// do { /// x = phi (x0, x.next); //PhiX /// cnt = phi(cnt0, cnt.next); /// /// cnt.next = cnt + 1; /// ... /// x.next = x >> 1; // DefX /// ... /// } while(x.next != 0); /// /// loop-exit: /// \endcode static bool detectShiftUntilZeroIdiom(Loop *CurLoop, const DataLayout &DL, Intrinsic::ID &IntrinID, Value *&InitX, Instruction *&CntInst, PHINode *&CntPhi, Instruction *&DefX) { BasicBlock *LoopEntry; Value *VarX = nullptr; DefX = nullptr; CntInst = nullptr; CntPhi = nullptr; LoopEntry = *(CurLoop->block_begin()); // step 1: Check if the loop-back branch is in desirable form. if (Value *T = matchCondition( dyn_cast(LoopEntry->getTerminator()), LoopEntry)) DefX = dyn_cast(T); else return false; // step 2: detect instructions corresponding to "x.next = x >> 1 or x << 1" if (!DefX || !DefX->isShift()) return false; IntrinID = DefX->getOpcode() == Instruction::Shl ? Intrinsic::cttz : Intrinsic::ctlz; ConstantInt *Shft = dyn_cast(DefX->getOperand(1)); if (!Shft || !Shft->isOne()) return false; VarX = DefX->getOperand(0); // step 3: Check the recurrence of variable X PHINode *PhiX = getRecurrenceVar(VarX, DefX, LoopEntry); if (!PhiX) return false; InitX = PhiX->getIncomingValueForBlock(CurLoop->getLoopPreheader()); // Make sure the initial value can't be negative otherwise the ashr in the // loop might never reach zero which would make the loop infinite. if (DefX->getOpcode() == Instruction::AShr && !isKnownNonNegative(InitX, DL)) return false; // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1 // or cnt.next = cnt + -1. // TODO: We can skip the step. If loop trip count is known (CTLZ), // then all uses of "cnt.next" could be optimized to the trip count // plus "cnt0". Currently it is not optimized. // This step could be used to detect POPCNT instruction: // cnt.next = cnt + (x.next & 1) for (Instruction &Inst : llvm::make_range( LoopEntry->getFirstNonPHI()->getIterator(), LoopEntry->end())) { if (Inst.getOpcode() != Instruction::Add) continue; ConstantInt *Inc = dyn_cast(Inst.getOperand(1)); if (!Inc || (!Inc->isOne() && !Inc->isMinusOne())) continue; PHINode *Phi = getRecurrenceVar(Inst.getOperand(0), &Inst, LoopEntry); if (!Phi) continue; CntInst = &Inst; CntPhi = Phi; break; } if (!CntInst) return false; return true; } /// Recognize CTLZ or CTTZ idiom in a non-countable loop and convert the loop /// to countable (with CTLZ / CTTZ trip count). If CTLZ / CTTZ inserted as a new /// trip count returns true; otherwise, returns false. bool LoopIdiomRecognize::recognizeAndInsertFFS() { // Give up if the loop has multiple blocks or multiple backedges. if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) return false; Intrinsic::ID IntrinID; Value *InitX; Instruction *DefX = nullptr; PHINode *CntPhi = nullptr; Instruction *CntInst = nullptr; // Help decide if transformation is profitable. For ShiftUntilZero idiom, // this is always 6. size_t IdiomCanonicalSize = 6; if (!detectShiftUntilZeroIdiom(CurLoop, *DL, IntrinID, InitX, CntInst, CntPhi, DefX)) return false; bool IsCntPhiUsedOutsideLoop = false; for (User *U : CntPhi->users()) if (!CurLoop->contains(cast(U))) { IsCntPhiUsedOutsideLoop = true; break; } bool IsCntInstUsedOutsideLoop = false; for (User *U : CntInst->users()) if (!CurLoop->contains(cast(U))) { IsCntInstUsedOutsideLoop = true; break; } // If both CntInst and CntPhi are used outside the loop the profitability // is questionable. if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop) return false; // For some CPUs result of CTLZ(X) intrinsic is undefined // when X is 0. If we can not guarantee X != 0, we need to check this // when expand. bool ZeroCheck = false; // It is safe to assume Preheader exist as it was checked in // parent function RunOnLoop. BasicBlock *PH = CurLoop->getLoopPreheader(); // If we are using the count instruction outside the loop, make sure we // have a zero check as a precondition. Without the check the loop would run // one iteration for before any check of the input value. This means 0 and 1 // would have identical behavior in the original loop and thus if (!IsCntPhiUsedOutsideLoop) { auto *PreCondBB = PH->getSinglePredecessor(); if (!PreCondBB) return false; auto *PreCondBI = dyn_cast(PreCondBB->getTerminator()); if (!PreCondBI) return false; if (matchCondition(PreCondBI, PH) != InitX) return false; ZeroCheck = true; } // Check if CTLZ / CTTZ intrinsic is profitable. Assume it is always // profitable if we delete the loop. // the loop has only 6 instructions: // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ] // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ] // %shr = ashr %n.addr.0, 1 // %tobool = icmp eq %shr, 0 // %inc = add nsw %i.0, 1 // br i1 %tobool const Value *Args[] = {InitX, ConstantInt::getBool(InitX->getContext(), ZeroCheck)}; // @llvm.dbg doesn't count as they have no semantic effect. auto InstWithoutDebugIt = CurLoop->getHeader()->instructionsWithoutDebug(); uint32_t HeaderSize = std::distance(InstWithoutDebugIt.begin(), InstWithoutDebugIt.end()); IntrinsicCostAttributes Attrs(IntrinID, InitX->getType(), Args); InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, TargetTransformInfo::TCK_SizeAndLatency); if (HeaderSize != IdiomCanonicalSize && Cost > TargetTransformInfo::TCC_Basic) return false; transformLoopToCountable(IntrinID, PH, CntInst, CntPhi, InitX, DefX, DefX->getDebugLoc(), ZeroCheck, IsCntPhiUsedOutsideLoop); return true; } /// Recognizes a population count idiom in a non-countable loop. /// /// If detected, transforms the relevant code to issue the popcount intrinsic /// function call, and returns true; otherwise, returns false. bool LoopIdiomRecognize::recognizePopcount() { if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware) return false; // Counting population are usually conducted by few arithmetic instructions. // Such instructions can be easily "absorbed" by vacant slots in a // non-compact loop. Therefore, recognizing popcount idiom only makes sense // in a compact loop. // Give up if the loop has multiple blocks or multiple backedges. if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) return false; BasicBlock *LoopBody = *(CurLoop->block_begin()); if (LoopBody->size() >= 20) { // The loop is too big, bail out. return false; } // It should have a preheader containing nothing but an unconditional branch. BasicBlock *PH = CurLoop->getLoopPreheader(); if (!PH || &PH->front() != PH->getTerminator()) return false; auto *EntryBI = dyn_cast(PH->getTerminator()); if (!EntryBI || EntryBI->isConditional()) return false; // It should have a precondition block where the generated popcount intrinsic // function can be inserted. auto *PreCondBB = PH->getSinglePredecessor(); if (!PreCondBB) return false; auto *PreCondBI = dyn_cast(PreCondBB->getTerminator()); if (!PreCondBI || PreCondBI->isUnconditional()) return false; Instruction *CntInst; PHINode *CntPhi; Value *Val; if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val)) return false; transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val); return true; } static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val, const DebugLoc &DL) { Value *Ops[] = {Val}; Type *Tys[] = {Val->getType()}; Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); Function *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys); CallInst *CI = IRBuilder.CreateCall(Func, Ops); CI->setDebugLoc(DL); return CI; } static CallInst *createFFSIntrinsic(IRBuilder<> &IRBuilder, Value *Val, const DebugLoc &DL, bool ZeroCheck, Intrinsic::ID IID) { Value *Ops[] = {Val, IRBuilder.getInt1(ZeroCheck)}; Type *Tys[] = {Val->getType()}; Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); Function *Func = Intrinsic::getDeclaration(M, IID, Tys); CallInst *CI = IRBuilder.CreateCall(Func, Ops); CI->setDebugLoc(DL); return CI; } /// Transform the following loop (Using CTLZ, CTTZ is similar): /// loop: /// CntPhi = PHI [Cnt0, CntInst] /// PhiX = PHI [InitX, DefX] /// CntInst = CntPhi + 1 /// DefX = PhiX >> 1 /// LOOP_BODY /// Br: loop if (DefX != 0) /// Use(CntPhi) or Use(CntInst) /// /// Into: /// If CntPhi used outside the loop: /// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1) /// Count = CountPrev + 1 /// else /// Count = BitWidth(InitX) - CTLZ(InitX) /// loop: /// CntPhi = PHI [Cnt0, CntInst] /// PhiX = PHI [InitX, DefX] /// PhiCount = PHI [Count, Dec] /// CntInst = CntPhi + 1 /// DefX = PhiX >> 1 /// Dec = PhiCount - 1 /// LOOP_BODY /// Br: loop if (Dec != 0) /// Use(CountPrev + Cnt0) // Use(CntPhi) /// or /// Use(Count + Cnt0) // Use(CntInst) /// /// If LOOP_BODY is empty the loop will be deleted. /// If CntInst and DefX are not used in LOOP_BODY they will be removed. void LoopIdiomRecognize::transformLoopToCountable( Intrinsic::ID IntrinID, BasicBlock *Preheader, Instruction *CntInst, PHINode *CntPhi, Value *InitX, Instruction *DefX, const DebugLoc &DL, bool ZeroCheck, bool IsCntPhiUsedOutsideLoop) { BranchInst *PreheaderBr = cast(Preheader->getTerminator()); // Step 1: Insert the CTLZ/CTTZ instruction at the end of the preheader block IRBuilder<> Builder(PreheaderBr); Builder.SetCurrentDebugLocation(DL); // If there are no uses of CntPhi crate: // Count = BitWidth - CTLZ(InitX); // NewCount = Count; // If there are uses of CntPhi create: // NewCount = BitWidth - CTLZ(InitX >> 1); // Count = NewCount + 1; Value *InitXNext; if (IsCntPhiUsedOutsideLoop) { if (DefX->getOpcode() == Instruction::AShr) InitXNext = Builder.CreateAShr(InitX, 1); else if (DefX->getOpcode() == Instruction::LShr) InitXNext = Builder.CreateLShr(InitX, 1); else if (DefX->getOpcode() == Instruction::Shl) // cttz InitXNext = Builder.CreateShl(InitX, 1); else llvm_unreachable("Unexpected opcode!"); } else InitXNext = InitX; Value *Count = createFFSIntrinsic(Builder, InitXNext, DL, ZeroCheck, IntrinID); Type *CountTy = Count->getType(); Count = Builder.CreateSub( ConstantInt::get(CountTy, CountTy->getIntegerBitWidth()), Count); Value *NewCount = Count; if (IsCntPhiUsedOutsideLoop) Count = Builder.CreateAdd(Count, ConstantInt::get(CountTy, 1)); NewCount = Builder.CreateZExtOrTrunc(NewCount, CntInst->getType()); Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader); if (cast(CntInst->getOperand(1))->isOne()) { // If the counter was being incremented in the loop, add NewCount to the // counter's initial value, but only if the initial value is not zero. ConstantInt *InitConst = dyn_cast(CntInitVal); if (!InitConst || !InitConst->isZero()) NewCount = Builder.CreateAdd(NewCount, CntInitVal); } else { // If the count was being decremented in the loop, subtract NewCount from // the counter's initial value. NewCount = Builder.CreateSub(CntInitVal, NewCount); } // Step 2: Insert new IV and loop condition: // loop: // ... // PhiCount = PHI [Count, Dec] // ... // Dec = PhiCount - 1 // ... // Br: loop if (Dec != 0) BasicBlock *Body = *(CurLoop->block_begin()); auto *LbBr = cast(Body->getTerminator()); ICmpInst *LbCond = cast(LbBr->getCondition()); PHINode *TcPhi = PHINode::Create(CountTy, 2, "tcphi", &Body->front()); Builder.SetInsertPoint(LbCond); Instruction *TcDec = cast(Builder.CreateSub( TcPhi, ConstantInt::get(CountTy, 1), "tcdec", false, true)); TcPhi->addIncoming(Count, Preheader); TcPhi->addIncoming(TcDec, Body); CmpInst::Predicate Pred = (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ; LbCond->setPredicate(Pred); LbCond->setOperand(0, TcDec); LbCond->setOperand(1, ConstantInt::get(CountTy, 0)); // Step 3: All the references to the original counter outside // the loop are replaced with the NewCount if (IsCntPhiUsedOutsideLoop) CntPhi->replaceUsesOutsideBlock(NewCount, Body); else CntInst->replaceUsesOutsideBlock(NewCount, Body); // step 4: Forget the "non-computable" trip-count SCEV associated with the // loop. The loop would otherwise not be deleted even if it becomes empty. SE->forgetLoop(CurLoop); } void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst, PHINode *CntPhi, Value *Var) { BasicBlock *PreHead = CurLoop->getLoopPreheader(); auto *PreCondBr = cast(PreCondBB->getTerminator()); const DebugLoc &DL = CntInst->getDebugLoc(); // Assuming before transformation, the loop is following: // if (x) // the precondition // do { cnt++; x &= x - 1; } while(x); // Step 1: Insert the ctpop instruction at the end of the precondition block IRBuilder<> Builder(PreCondBr); Value *PopCnt, *PopCntZext, *NewCount, *TripCnt; { PopCnt = createPopcntIntrinsic(Builder, Var, DL); NewCount = PopCntZext = Builder.CreateZExtOrTrunc(PopCnt, cast(CntPhi->getType())); if (NewCount != PopCnt) (cast(NewCount))->setDebugLoc(DL); // TripCnt is exactly the number of iterations the loop has TripCnt = NewCount; // If the population counter's initial value is not zero, insert Add Inst. Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead); ConstantInt *InitConst = dyn_cast(CntInitVal); if (!InitConst || !InitConst->isZero()) { NewCount = Builder.CreateAdd(NewCount, CntInitVal); (cast(NewCount))->setDebugLoc(DL); } } // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to // "if (NewCount == 0) loop-exit". Without this change, the intrinsic // function would be partial dead code, and downstream passes will drag // it back from the precondition block to the preheader. { ICmpInst *PreCond = cast(PreCondBr->getCondition()); Value *Opnd0 = PopCntZext; Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0); if (PreCond->getOperand(0) != Var) std::swap(Opnd0, Opnd1); ICmpInst *NewPreCond = cast( Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1)); PreCondBr->setCondition(NewPreCond); RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI); } // Step 3: Note that the population count is exactly the trip count of the // loop in question, which enable us to convert the loop from noncountable // loop into a countable one. The benefit is twofold: // // - If the loop only counts population, the entire loop becomes dead after // the transformation. It is a lot easier to prove a countable loop dead // than to prove a noncountable one. (In some C dialects, an infinite loop // isn't dead even if it computes nothing useful. In general, DCE needs // to prove a noncountable loop finite before safely delete it.) // // - If the loop also performs something else, it remains alive. // Since it is transformed to countable form, it can be aggressively // optimized by some optimizations which are in general not applicable // to a noncountable loop. // // After this step, this loop (conceptually) would look like following: // newcnt = __builtin_ctpop(x); // t = newcnt; // if (x) // do { cnt++; x &= x-1; t--) } while (t > 0); BasicBlock *Body = *(CurLoop->block_begin()); { auto *LbBr = cast(Body->getTerminator()); ICmpInst *LbCond = cast(LbBr->getCondition()); Type *Ty = TripCnt->getType(); PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front()); Builder.SetInsertPoint(LbCond); Instruction *TcDec = cast( Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1), "tcdec", false, true)); TcPhi->addIncoming(TripCnt, PreHead); TcPhi->addIncoming(TcDec, Body); CmpInst::Predicate Pred = (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE; LbCond->setPredicate(Pred); LbCond->setOperand(0, TcDec); LbCond->setOperand(1, ConstantInt::get(Ty, 0)); } // Step 4: All the references to the original population counter outside // the loop are replaced with the NewCount -- the value returned from // __builtin_ctpop(). CntInst->replaceUsesOutsideBlock(NewCount, Body); // step 5: Forget the "non-computable" trip-count SCEV associated with the // loop. The loop would otherwise not be deleted even if it becomes empty. SE->forgetLoop(CurLoop); } /// Match loop-invariant value. template struct match_LoopInvariant { SubPattern_t SubPattern; const Loop *L; match_LoopInvariant(const SubPattern_t &SP, const Loop *L) : SubPattern(SP), L(L) {} template bool match(ITy *V) { return L->isLoopInvariant(V) && SubPattern.match(V); } }; /// Matches if the value is loop-invariant. template inline match_LoopInvariant m_LoopInvariant(const Ty &M, const Loop *L) { return match_LoopInvariant(M, L); } /// Return true if the idiom is detected in the loop. /// /// The core idiom we are trying to detect is: /// \code /// entry: /// <...> /// %bitmask = shl i32 1, %bitpos /// br label %loop /// /// loop: /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ] /// %x.curr.bitmasked = and i32 %x.curr, %bitmask /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0 /// %x.next = shl i32 %x.curr, 1 /// <...> /// br i1 %x.curr.isbitunset, label %loop, label %end /// /// end: /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> /// %x.next.res = phi i32 [ %x.next, %loop ] <...> /// <...> /// \endcode static bool detectShiftUntilBitTestIdiom(Loop *CurLoop, Value *&BaseX, Value *&BitMask, Value *&BitPos, Value *&CurrX, Instruction *&NextX) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Performing shift-until-bittest idiom detection.\n"); // Give up if the loop has multiple blocks or multiple backedges. if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n"); return false; } BasicBlock *LoopHeaderBB = CurLoop->getHeader(); BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); assert(LoopPreheaderBB && "There is always a loop preheader."); using namespace PatternMatch; // Step 1: Check if the loop backedge is in desirable form. ICmpInst::Predicate Pred; Value *CmpLHS, *CmpRHS; BasicBlock *TrueBB, *FalseBB; if (!match(LoopHeaderBB->getTerminator(), m_Br(m_ICmp(Pred, m_Value(CmpLHS), m_Value(CmpRHS)), m_BasicBlock(TrueBB), m_BasicBlock(FalseBB)))) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n"); return false; } // Step 2: Check if the backedge's condition is in desirable form. auto MatchVariableBitMask = [&]() { return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) && match(CmpLHS, m_c_And(m_Value(CurrX), m_CombineAnd( m_Value(BitMask), m_LoopInvariant(m_Shl(m_One(), m_Value(BitPos)), CurLoop)))); }; auto MatchConstantBitMask = [&]() { return ICmpInst::isEquality(Pred) && match(CmpRHS, m_Zero()) && match(CmpLHS, m_And(m_Value(CurrX), m_CombineAnd(m_Value(BitMask), m_Power2()))) && (BitPos = ConstantExpr::getExactLogBase2(cast(BitMask))); }; auto MatchDecomposableConstantBitMask = [&]() { APInt Mask; return llvm::decomposeBitTestICmp(CmpLHS, CmpRHS, Pred, CurrX, Mask) && ICmpInst::isEquality(Pred) && Mask.isPowerOf2() && (BitMask = ConstantInt::get(CurrX->getType(), Mask)) && (BitPos = ConstantInt::get(CurrX->getType(), Mask.logBase2())); }; if (!MatchVariableBitMask() && !MatchConstantBitMask() && !MatchDecomposableConstantBitMask()) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge comparison.\n"); return false; } // Step 3: Check if the recurrence is in desirable form. auto *CurrXPN = dyn_cast(CurrX); if (!CurrXPN || CurrXPN->getParent() != LoopHeaderBB) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n"); return false; } BaseX = CurrXPN->getIncomingValueForBlock(LoopPreheaderBB); NextX = dyn_cast(CurrXPN->getIncomingValueForBlock(LoopHeaderBB)); assert(CurLoop->isLoopInvariant(BaseX) && "Expected BaseX to be avaliable in the preheader!"); if (!NextX || !match(NextX, m_Shl(m_Specific(CurrX), m_One()))) { // FIXME: support right-shift? LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n"); return false; } // Step 4: Check if the backedge's destinations are in desirable form. assert(ICmpInst::isEquality(Pred) && "Should only get equality predicates here."); // cmp-br is commutative, so canonicalize to a single variant. if (Pred != ICmpInst::Predicate::ICMP_EQ) { Pred = ICmpInst::getInversePredicate(Pred); std::swap(TrueBB, FalseBB); } // We expect to exit loop when comparison yields false, // so when it yields true we should branch back to loop header. if (TrueBB != LoopHeaderBB) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n"); return false; } // Okay, idiom checks out. return true; } /// Look for the following loop: /// \code /// entry: /// <...> /// %bitmask = shl i32 1, %bitpos /// br label %loop /// /// loop: /// %x.curr = phi i32 [ %x, %entry ], [ %x.next, %loop ] /// %x.curr.bitmasked = and i32 %x.curr, %bitmask /// %x.curr.isbitunset = icmp eq i32 %x.curr.bitmasked, 0 /// %x.next = shl i32 %x.curr, 1 /// <...> /// br i1 %x.curr.isbitunset, label %loop, label %end /// /// end: /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> /// %x.next.res = phi i32 [ %x.next, %loop ] <...> /// <...> /// \endcode /// /// And transform it into: /// \code /// entry: /// %bitmask = shl i32 1, %bitpos /// %lowbitmask = add i32 %bitmask, -1 /// %mask = or i32 %lowbitmask, %bitmask /// %x.masked = and i32 %x, %mask /// %x.masked.numleadingzeros = call i32 @llvm.ctlz.i32(i32 %x.masked, /// i1 true) /// %x.masked.numactivebits = sub i32 32, %x.masked.numleadingzeros /// %x.masked.leadingonepos = add i32 %x.masked.numactivebits, -1 /// %backedgetakencount = sub i32 %bitpos, %x.masked.leadingonepos /// %tripcount = add i32 %backedgetakencount, 1 /// %x.curr = shl i32 %x, %backedgetakencount /// %x.next = shl i32 %x, %tripcount /// br label %loop /// /// loop: /// %loop.iv = phi i32 [ 0, %entry ], [ %loop.iv.next, %loop ] /// %loop.iv.next = add nuw i32 %loop.iv, 1 /// %loop.ivcheck = icmp eq i32 %loop.iv.next, %tripcount /// <...> /// br i1 %loop.ivcheck, label %end, label %loop /// /// end: /// %x.curr.res = phi i32 [ %x.curr, %loop ] <...> /// %x.next.res = phi i32 [ %x.next, %loop ] <...> /// <...> /// \endcode bool LoopIdiomRecognize::recognizeShiftUntilBitTest() { bool MadeChange = false; Value *X, *BitMask, *BitPos, *XCurr; Instruction *XNext; if (!detectShiftUntilBitTestIdiom(CurLoop, X, BitMask, BitPos, XCurr, XNext)) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom detection failed.\n"); return MadeChange; } LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom detected!\n"); // Ok, it is the idiom we were looking for, we *could* transform this loop, // but is it profitable to transform? BasicBlock *LoopHeaderBB = CurLoop->getHeader(); BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); assert(LoopPreheaderBB && "There is always a loop preheader."); BasicBlock *SuccessorBB = CurLoop->getExitBlock(); assert(SuccessorBB && "There is only a single successor."); IRBuilder<> Builder(LoopPreheaderBB->getTerminator()); Builder.SetCurrentDebugLocation(cast(XCurr)->getDebugLoc()); Intrinsic::ID IntrID = Intrinsic::ctlz; Type *Ty = X->getType(); unsigned Bitwidth = Ty->getScalarSizeInBits(); TargetTransformInfo::TargetCostKind CostKind = TargetTransformInfo::TCK_SizeAndLatency; // The rewrite is considered to be unprofitable iff and only iff the // intrinsic/shift we'll use are not cheap. Note that we are okay with *just* // making the loop countable, even if nothing else changes. IntrinsicCostAttributes Attrs( IntrID, Ty, {PoisonValue::get(Ty), /*is_zero_poison=*/Builder.getTrue()}); InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind); if (Cost > TargetTransformInfo::TCC_Basic) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Intrinsic is too costly, not beneficial\n"); return MadeChange; } if (TTI->getArithmeticInstrCost(Instruction::Shl, Ty, CostKind) > TargetTransformInfo::TCC_Basic) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Shift is too costly, not beneficial\n"); return MadeChange; } // Ok, transform appears worthwhile. MadeChange = true; if (!isGuaranteedNotToBeUndefOrPoison(BitPos)) { // BitMask may be computed from BitPos, Freeze BitPos so we can increase // it's use count. Instruction *InsertPt = nullptr; if (auto *BitPosI = dyn_cast(BitPos)) InsertPt = BitPosI->getInsertionPointAfterDef(); else InsertPt = &*DT->getRoot()->getFirstNonPHIOrDbgOrAlloca(); if (!InsertPt) return false; FreezeInst *BitPosFrozen = new FreezeInst(BitPos, BitPos->getName() + ".fr", InsertPt); BitPos->replaceUsesWithIf(BitPosFrozen, [BitPosFrozen](Use &U) { return U.getUser() != BitPosFrozen; }); BitPos = BitPosFrozen; } // Step 1: Compute the loop trip count. Value *LowBitMask = Builder.CreateAdd(BitMask, Constant::getAllOnesValue(Ty), BitPos->getName() + ".lowbitmask"); Value *Mask = Builder.CreateOr(LowBitMask, BitMask, BitPos->getName() + ".mask"); Value *XMasked = Builder.CreateAnd(X, Mask, X->getName() + ".masked"); CallInst *XMaskedNumLeadingZeros = Builder.CreateIntrinsic( IntrID, Ty, {XMasked, /*is_zero_poison=*/Builder.getTrue()}, /*FMFSource=*/nullptr, XMasked->getName() + ".numleadingzeros"); Value *XMaskedNumActiveBits = Builder.CreateSub( ConstantInt::get(Ty, Ty->getScalarSizeInBits()), XMaskedNumLeadingZeros, XMasked->getName() + ".numactivebits", /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); Value *XMaskedLeadingOnePos = Builder.CreateAdd(XMaskedNumActiveBits, Constant::getAllOnesValue(Ty), XMasked->getName() + ".leadingonepos", /*HasNUW=*/false, /*HasNSW=*/Bitwidth > 2); Value *LoopBackedgeTakenCount = Builder.CreateSub( BitPos, XMaskedLeadingOnePos, CurLoop->getName() + ".backedgetakencount", /*HasNUW=*/true, /*HasNSW=*/true); // We know loop's backedge-taken count, but what's loop's trip count? // Note that while NUW is always safe, while NSW is only for bitwidths != 2. Value *LoopTripCount = Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1), CurLoop->getName() + ".tripcount", /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); // Step 2: Compute the recurrence's final value without a loop. // NewX is always safe to compute, because `LoopBackedgeTakenCount` // will always be smaller than `bitwidth(X)`, i.e. we never get poison. Value *NewX = Builder.CreateShl(X, LoopBackedgeTakenCount); NewX->takeName(XCurr); if (auto *I = dyn_cast(NewX)) I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true); Value *NewXNext; // Rewriting XNext is more complicated, however, because `X << LoopTripCount` // will be poison iff `LoopTripCount == bitwidth(X)` (which will happen // iff `BitPos` is `bitwidth(x) - 1` and `X` is `1`). So unless we know // that isn't the case, we'll need to emit an alternative, safe IR. if (XNext->hasNoSignedWrap() || XNext->hasNoUnsignedWrap() || PatternMatch::match( BitPos, PatternMatch::m_SpecificInt_ICMP( ICmpInst::ICMP_NE, APInt(Ty->getScalarSizeInBits(), Ty->getScalarSizeInBits() - 1)))) NewXNext = Builder.CreateShl(X, LoopTripCount); else { // Otherwise, just additionally shift by one. It's the smallest solution, // alternatively, we could check that NewX is INT_MIN (or BitPos is ) // and select 0 instead. NewXNext = Builder.CreateShl(NewX, ConstantInt::get(Ty, 1)); } NewXNext->takeName(XNext); if (auto *I = dyn_cast(NewXNext)) I->copyIRFlags(XNext, /*IncludeWrapFlags=*/true); // Step 3: Adjust the successor basic block to recieve the computed // recurrence's final value instead of the recurrence itself. XCurr->replaceUsesOutsideBlock(NewX, LoopHeaderBB); XNext->replaceUsesOutsideBlock(NewXNext, LoopHeaderBB); // Step 4: Rewrite the loop into a countable form, with canonical IV. // The new canonical induction variable. Builder.SetInsertPoint(&LoopHeaderBB->front()); auto *IV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv"); // The induction itself. // Note that while NUW is always safe, while NSW is only for bitwidths != 2. Builder.SetInsertPoint(LoopHeaderBB->getTerminator()); auto *IVNext = Builder.CreateAdd(IV, ConstantInt::get(Ty, 1), IV->getName() + ".next", /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); // The loop trip count check. auto *IVCheck = Builder.CreateICmpEQ(IVNext, LoopTripCount, CurLoop->getName() + ".ivcheck"); Builder.CreateCondBr(IVCheck, SuccessorBB, LoopHeaderBB); LoopHeaderBB->getTerminator()->eraseFromParent(); // Populate the IV PHI. IV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB); IV->addIncoming(IVNext, LoopHeaderBB); // Step 5: Forget the "non-computable" trip-count SCEV associated with the // loop. The loop would otherwise not be deleted even if it becomes empty. SE->forgetLoop(CurLoop); // Other passes will take care of actually deleting the loop if possible. LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-bittest idiom optimized!\n"); ++NumShiftUntilBitTest; return MadeChange; } /// Return true if the idiom is detected in the loop. /// /// The core idiom we are trying to detect is: /// \code /// entry: /// <...> /// %start = <...> /// %extraoffset = <...> /// <...> /// br label %for.cond /// /// loop: /// %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ] /// %nbits = add nsw i8 %iv, %extraoffset /// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits /// %val.shifted.iszero = icmp eq i8 %val.shifted, 0 /// %iv.next = add i8 %iv, 1 /// <...> /// br i1 %val.shifted.iszero, label %end, label %loop /// /// end: /// %iv.res = phi i8 [ %iv, %loop ] <...> /// %nbits.res = phi i8 [ %nbits, %loop ] <...> /// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...> /// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...> /// %iv.next.res = phi i8 [ %iv.next, %loop ] <...> /// <...> /// \endcode static bool detectShiftUntilZeroIdiom(Loop *CurLoop, ScalarEvolution *SE, Instruction *&ValShiftedIsZero, Intrinsic::ID &IntrinID, Instruction *&IV, Value *&Start, Value *&Val, const SCEV *&ExtraOffsetExpr, bool &InvertedCond) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Performing shift-until-zero idiom detection.\n"); // Give up if the loop has multiple blocks or multiple backedges. if (CurLoop->getNumBlocks() != 1 || CurLoop->getNumBackEdges() != 1) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad block/backedge count.\n"); return false; } Instruction *ValShifted, *NBits, *IVNext; Value *ExtraOffset; BasicBlock *LoopHeaderBB = CurLoop->getHeader(); BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); assert(LoopPreheaderBB && "There is always a loop preheader."); using namespace PatternMatch; // Step 1: Check if the loop backedge, condition is in desirable form. ICmpInst::Predicate Pred; BasicBlock *TrueBB, *FalseBB; if (!match(LoopHeaderBB->getTerminator(), m_Br(m_Instruction(ValShiftedIsZero), m_BasicBlock(TrueBB), m_BasicBlock(FalseBB))) || !match(ValShiftedIsZero, m_ICmp(Pred, m_Instruction(ValShifted), m_Zero())) || !ICmpInst::isEquality(Pred)) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge structure.\n"); return false; } // Step 2: Check if the comparison's operand is in desirable form. // FIXME: Val could be a one-input PHI node, which we should look past. if (!match(ValShifted, m_Shift(m_LoopInvariant(m_Value(Val), CurLoop), m_Instruction(NBits)))) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad comparisons value computation.\n"); return false; } IntrinID = ValShifted->getOpcode() == Instruction::Shl ? Intrinsic::cttz : Intrinsic::ctlz; // Step 3: Check if the shift amount is in desirable form. if (match(NBits, m_c_Add(m_Instruction(IV), m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) && (NBits->hasNoSignedWrap() || NBits->hasNoUnsignedWrap())) ExtraOffsetExpr = SE->getNegativeSCEV(SE->getSCEV(ExtraOffset)); else if (match(NBits, m_Sub(m_Instruction(IV), m_LoopInvariant(m_Value(ExtraOffset), CurLoop))) && NBits->hasNoSignedWrap()) ExtraOffsetExpr = SE->getSCEV(ExtraOffset); else { IV = NBits; ExtraOffsetExpr = SE->getZero(NBits->getType()); } // Step 4: Check if the recurrence is in desirable form. auto *IVPN = dyn_cast(IV); if (!IVPN || IVPN->getParent() != LoopHeaderBB) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Not an expected PHI node.\n"); return false; } Start = IVPN->getIncomingValueForBlock(LoopPreheaderBB); IVNext = dyn_cast(IVPN->getIncomingValueForBlock(LoopHeaderBB)); if (!IVNext || !match(IVNext, m_Add(m_Specific(IVPN), m_One()))) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad recurrence.\n"); return false; } // Step 4: Check if the backedge's destinations are in desirable form. assert(ICmpInst::isEquality(Pred) && "Should only get equality predicates here."); // cmp-br is commutative, so canonicalize to a single variant. InvertedCond = Pred != ICmpInst::Predicate::ICMP_EQ; if (InvertedCond) { Pred = ICmpInst::getInversePredicate(Pred); std::swap(TrueBB, FalseBB); } // We expect to exit loop when comparison yields true, // so when it yields false we should branch back to loop header. if (FalseBB != LoopHeaderBB) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Bad backedge flow.\n"); return false; } // The new, countable, loop will certainly only run a known number of // iterations, It won't be infinite. But the old loop might be infinite // under certain conditions. For logical shifts, the value will become zero // after at most bitwidth(%Val) loop iterations. However, for arithmetic // right-shift, iff the sign bit was set, the value will never become zero, // and the loop may never finish. if (ValShifted->getOpcode() == Instruction::AShr && !isMustProgress(CurLoop) && !SE->isKnownNonNegative(SE->getSCEV(Val))) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Can not prove the loop is finite.\n"); return false; } // Okay, idiom checks out. return true; } /// Look for the following loop: /// \code /// entry: /// <...> /// %start = <...> /// %extraoffset = <...> /// <...> /// br label %for.cond /// /// loop: /// %iv = phi i8 [ %start, %entry ], [ %iv.next, %for.cond ] /// %nbits = add nsw i8 %iv, %extraoffset /// %val.shifted = {{l,a}shr,shl} i8 %val, %nbits /// %val.shifted.iszero = icmp eq i8 %val.shifted, 0 /// %iv.next = add i8 %iv, 1 /// <...> /// br i1 %val.shifted.iszero, label %end, label %loop /// /// end: /// %iv.res = phi i8 [ %iv, %loop ] <...> /// %nbits.res = phi i8 [ %nbits, %loop ] <...> /// %val.shifted.res = phi i8 [ %val.shifted, %loop ] <...> /// %val.shifted.iszero.res = phi i1 [ %val.shifted.iszero, %loop ] <...> /// %iv.next.res = phi i8 [ %iv.next, %loop ] <...> /// <...> /// \endcode /// /// And transform it into: /// \code /// entry: /// <...> /// %start = <...> /// %extraoffset = <...> /// <...> /// %val.numleadingzeros = call i8 @llvm.ct{l,t}z.i8(i8 %val, i1 0) /// %val.numactivebits = sub i8 8, %val.numleadingzeros /// %extraoffset.neg = sub i8 0, %extraoffset /// %tmp = add i8 %val.numactivebits, %extraoffset.neg /// %iv.final = call i8 @llvm.smax.i8(i8 %tmp, i8 %start) /// %loop.tripcount = sub i8 %iv.final, %start /// br label %loop /// /// loop: /// %loop.iv = phi i8 [ 0, %entry ], [ %loop.iv.next, %loop ] /// %loop.iv.next = add i8 %loop.iv, 1 /// %loop.ivcheck = icmp eq i8 %loop.iv.next, %loop.tripcount /// %iv = add i8 %loop.iv, %start /// <...> /// br i1 %loop.ivcheck, label %end, label %loop /// /// end: /// %iv.res = phi i8 [ %iv.final, %loop ] <...> /// <...> /// \endcode bool LoopIdiomRecognize::recognizeShiftUntilZero() { bool MadeChange = false; Instruction *ValShiftedIsZero; Intrinsic::ID IntrID; Instruction *IV; Value *Start, *Val; const SCEV *ExtraOffsetExpr; bool InvertedCond; if (!detectShiftUntilZeroIdiom(CurLoop, SE, ValShiftedIsZero, IntrID, IV, Start, Val, ExtraOffsetExpr, InvertedCond)) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom detection failed.\n"); return MadeChange; } LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom detected!\n"); // Ok, it is the idiom we were looking for, we *could* transform this loop, // but is it profitable to transform? BasicBlock *LoopHeaderBB = CurLoop->getHeader(); BasicBlock *LoopPreheaderBB = CurLoop->getLoopPreheader(); assert(LoopPreheaderBB && "There is always a loop preheader."); BasicBlock *SuccessorBB = CurLoop->getExitBlock(); assert(SuccessorBB && "There is only a single successor."); IRBuilder<> Builder(LoopPreheaderBB->getTerminator()); Builder.SetCurrentDebugLocation(IV->getDebugLoc()); Type *Ty = Val->getType(); unsigned Bitwidth = Ty->getScalarSizeInBits(); TargetTransformInfo::TargetCostKind CostKind = TargetTransformInfo::TCK_SizeAndLatency; // The rewrite is considered to be unprofitable iff and only iff the // intrinsic we'll use are not cheap. Note that we are okay with *just* // making the loop countable, even if nothing else changes. IntrinsicCostAttributes Attrs( IntrID, Ty, {PoisonValue::get(Ty), /*is_zero_poison=*/Builder.getFalse()}); InstructionCost Cost = TTI->getIntrinsicInstrCost(Attrs, CostKind); if (Cost > TargetTransformInfo::TCC_Basic) { LLVM_DEBUG(dbgs() << DEBUG_TYPE " Intrinsic is too costly, not beneficial\n"); return MadeChange; } // Ok, transform appears worthwhile. MadeChange = true; bool OffsetIsZero = false; if (auto *ExtraOffsetExprC = dyn_cast(ExtraOffsetExpr)) OffsetIsZero = ExtraOffsetExprC->isZero(); // Step 1: Compute the loop's final IV value / trip count. CallInst *ValNumLeadingZeros = Builder.CreateIntrinsic( IntrID, Ty, {Val, /*is_zero_poison=*/Builder.getFalse()}, /*FMFSource=*/nullptr, Val->getName() + ".numleadingzeros"); Value *ValNumActiveBits = Builder.CreateSub( ConstantInt::get(Ty, Ty->getScalarSizeInBits()), ValNumLeadingZeros, Val->getName() + ".numactivebits", /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); SCEVExpander Expander(*SE, *DL, "loop-idiom"); Expander.setInsertPoint(&*Builder.GetInsertPoint()); Value *ExtraOffset = Expander.expandCodeFor(ExtraOffsetExpr); Value *ValNumActiveBitsOffset = Builder.CreateAdd( ValNumActiveBits, ExtraOffset, ValNumActiveBits->getName() + ".offset", /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true); Value *IVFinal = Builder.CreateIntrinsic(Intrinsic::smax, {Ty}, {ValNumActiveBitsOffset, Start}, /*FMFSource=*/nullptr, "iv.final"); auto *LoopBackedgeTakenCount = cast(Builder.CreateSub( IVFinal, Start, CurLoop->getName() + ".backedgetakencount", /*HasNUW=*/OffsetIsZero, /*HasNSW=*/true)); // FIXME: or when the offset was `add nuw` // We know loop's backedge-taken count, but what's loop's trip count? Value *LoopTripCount = Builder.CreateAdd(LoopBackedgeTakenCount, ConstantInt::get(Ty, 1), CurLoop->getName() + ".tripcount", /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); // Step 2: Adjust the successor basic block to recieve the original // induction variable's final value instead of the orig. IV itself. IV->replaceUsesOutsideBlock(IVFinal, LoopHeaderBB); // Step 3: Rewrite the loop into a countable form, with canonical IV. // The new canonical induction variable. Builder.SetInsertPoint(&LoopHeaderBB->front()); auto *CIV = Builder.CreatePHI(Ty, 2, CurLoop->getName() + ".iv"); // The induction itself. Builder.SetInsertPoint(LoopHeaderBB->getFirstNonPHI()); auto *CIVNext = Builder.CreateAdd(CIV, ConstantInt::get(Ty, 1), CIV->getName() + ".next", /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); // The loop trip count check. auto *CIVCheck = Builder.CreateICmpEQ(CIVNext, LoopTripCount, CurLoop->getName() + ".ivcheck"); auto *NewIVCheck = CIVCheck; if (InvertedCond) { NewIVCheck = Builder.CreateNot(CIVCheck); NewIVCheck->takeName(ValShiftedIsZero); } // The original IV, but rebased to be an offset to the CIV. auto *IVDePHId = Builder.CreateAdd(CIV, Start, "", /*HasNUW=*/false, /*HasNSW=*/true); // FIXME: what about NUW? IVDePHId->takeName(IV); // The loop terminator. Builder.SetInsertPoint(LoopHeaderBB->getTerminator()); Builder.CreateCondBr(CIVCheck, SuccessorBB, LoopHeaderBB); LoopHeaderBB->getTerminator()->eraseFromParent(); // Populate the IV PHI. CIV->addIncoming(ConstantInt::get(Ty, 0), LoopPreheaderBB); CIV->addIncoming(CIVNext, LoopHeaderBB); // Step 4: Forget the "non-computable" trip-count SCEV associated with the // loop. The loop would otherwise not be deleted even if it becomes empty. SE->forgetLoop(CurLoop); // Step 5: Try to cleanup the loop's body somewhat. IV->replaceAllUsesWith(IVDePHId); IV->eraseFromParent(); ValShiftedIsZero->replaceAllUsesWith(NewIVCheck); ValShiftedIsZero->eraseFromParent(); // Other passes will take care of actually deleting the loop if possible. LLVM_DEBUG(dbgs() << DEBUG_TYPE " shift-until-zero idiom optimized!\n"); ++NumShiftUntilZero; return MadeChange; }