//===-- LoopIdiomRecognize.cpp - Loop idiom recognition -------------------===// // // The LLVM Compiler Infrastructure // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // // 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. // //===----------------------------------------------------------------------===// // // TODO List: // // Future loop memory idioms to recognize: // memcmp, memmove, strlen, etc. // Future floating point idioms to recognize in -ffast-math mode: // fpowi // Future integer operation idioms to recognize: // ctpop, ctlz, cttz // // 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. // // We should enhance the memset/memcpy recognition to handle multiple stores in // the loop. This would handle things like: // void foo(_Complex float *P) // for (i) { __real__(*P) = 0; __imag__(*P) = 0; } // // We should enhance this to handle negative strides through memory. // Alternatively (and perhaps better) we could rely on an earlier pass to force // forward iteration through memory, which is generally better for cache // behavior. Negative strides *do* happen for memset/memcpy loops. // // This could recognize common matrix multiplies and dot product idioms and // replace them with calls to BLAS (if linked in??). // //===----------------------------------------------------------------------===// #define DEBUG_TYPE "loop-idiom" #include "llvm/Transforms/Scalar.h" #include "llvm/IntrinsicInst.h" #include "llvm/Module.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/Analysis/LoopPass.h" #include "llvm/Analysis/ScalarEvolutionExpressions.h" #include "llvm/Analysis/ScalarEvolutionExpander.h" #include "llvm/Analysis/ValueTracking.h" #include "llvm/Target/TargetData.h" #include "llvm/Target/TargetLibraryInfo.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Support/Debug.h" #include "llvm/Support/IRBuilder.h" #include "llvm/Support/raw_ostream.h" #include "llvm/ADT/Statistic.h" using namespace llvm; STATISTIC(NumMemSet, "Number of memset's formed from loop stores"); STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores"); namespace { class LoopIdiomRecognize : public LoopPass { Loop *CurLoop; const TargetData *TD; DominatorTree *DT; ScalarEvolution *SE; TargetLibraryInfo *TLI; public: static char ID; explicit LoopIdiomRecognize() : LoopPass(ID) { initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry()); } bool runOnLoop(Loop *L, LPPassManager &LPM); bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount, SmallVectorImpl &ExitBlocks); bool processLoopStore(StoreInst *SI, const SCEV *BECount); bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount); bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize, unsigned StoreAlignment, Value *SplatValue, Instruction *TheStore, const SCEVAddRecExpr *Ev, const SCEV *BECount); bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize, const SCEVAddRecExpr *StoreEv, const SCEVAddRecExpr *LoadEv, const SCEV *BECount); /// This transformation requires natural loop information & requires that /// loop preheaders be inserted into the CFG. /// virtual void getAnalysisUsage(AnalysisUsage &AU) const { AU.addRequired(); AU.addPreserved(); AU.addRequiredID(LoopSimplifyID); AU.addPreservedID(LoopSimplifyID); AU.addRequiredID(LCSSAID); AU.addPreservedID(LCSSAID); AU.addRequired(); AU.addPreserved(); AU.addRequired(); AU.addPreserved(); AU.addPreserved(); AU.addRequired(); AU.addRequired(); } }; } char LoopIdiomRecognize::ID = 0; INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms", false, false) INITIALIZE_PASS_DEPENDENCY(LoopInfo) INITIALIZE_PASS_DEPENDENCY(DominatorTree) INITIALIZE_PASS_DEPENDENCY(LoopSimplify) INITIALIZE_PASS_DEPENDENCY(LCSSA) INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) INITIALIZE_AG_DEPENDENCY(AliasAnalysis) INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms", false, false) Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); } /// deleteDeadInstruction - Delete this instruction. Before we do, go through /// and zero out all the operands of this instruction. If any of them become /// dead, delete them and the computation tree that feeds them. /// static void deleteDeadInstruction(Instruction *I, ScalarEvolution &SE) { SmallVector NowDeadInsts; NowDeadInsts.push_back(I); // Before we touch this instruction, remove it from SE! do { Instruction *DeadInst = NowDeadInsts.pop_back_val(); // This instruction is dead, zap it, in stages. Start by removing it from // SCEV. SE.forgetValue(DeadInst); for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) { Value *Op = DeadInst->getOperand(op); DeadInst->setOperand(op, 0); // If this operand just became dead, add it to the NowDeadInsts list. if (!Op->use_empty()) continue; if (Instruction *OpI = dyn_cast(Op)) if (isInstructionTriviallyDead(OpI)) NowDeadInsts.push_back(OpI); } DeadInst->eraseFromParent(); } while (!NowDeadInsts.empty()); } /// deleteIfDeadInstruction - If the specified value is a dead instruction, /// delete it and any recursively used instructions. static void deleteIfDeadInstruction(Value *V, ScalarEvolution &SE) { if (Instruction *I = dyn_cast(V)) if (isInstructionTriviallyDead(I)) deleteDeadInstruction(I, SE); } bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) { CurLoop = L; // 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; // The trip count of the loop must be analyzable. SE = &getAnalysis(); if (!SE->hasLoopInvariantBackedgeTakenCount(L)) return false; const SCEV *BECount = SE->getBackedgeTakenCount(L); if (isa(BECount)) return false; // 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->getValue()->getValue() == 0) return false; // We require target data for now. TD = getAnalysisIfAvailable(); if (TD == 0) return false; DT = &getAnalysis(); LoopInfo &LI = getAnalysis(); TLI = &getAnalysis(); SmallVector ExitBlocks; CurLoop->getUniqueExitBlocks(ExitBlocks); DEBUG(dbgs() << "loop-idiom Scanning: F[" << L->getHeader()->getParent()->getName() << "] Loop %" << L->getHeader()->getName() << "\n"); bool MadeChange = false; // Scan all the blocks in the loop that are not in subloops. for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E; ++BI) { // Ignore blocks in subloops. if (LI.getLoopFor(*BI) != CurLoop) continue; MadeChange |= runOnLoopBlock(*BI, BECount, ExitBlocks); } return MadeChange; } /// 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 (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) if (!DT->dominates(BB, ExitBlocks[i])) return false; bool MadeChange = false; for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) { Instruction *Inst = I++; // Look for store instructions, which may be optimized to memset/memcpy. if (StoreInst *SI = dyn_cast(Inst)) { WeakVH InstPtr(I); if (!processLoopStore(SI, BECount)) continue; MadeChange = true; // If processing the store invalidated our iterator, start over from the // top of the block. if (InstPtr == 0) I = BB->begin(); continue; } // Look for memset instructions, which may be optimized to a larger memset. if (MemSetInst *MSI = dyn_cast(Inst)) { WeakVH InstPtr(I); if (!processLoopMemSet(MSI, BECount)) continue; MadeChange = true; // If processing the memset invalidated our iterator, start over from the // top of the block. if (InstPtr == 0) I = BB->begin(); continue; } } return MadeChange; } /// processLoopStore - See if this store can be promoted to a memset or memcpy. bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) { if (!SI->isSimple()) return false; Value *StoredVal = SI->getValueOperand(); Value *StorePtr = SI->getPointerOperand(); // Reject stores that are so large that they overflow an unsigned. uint64_t SizeInBits = TD->getTypeSizeInBits(StoredVal->getType()); if ((SizeInBits & 7) || (SizeInBits >> 32) != 0) return false; // 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 == 0 || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine()) return false; // 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. unsigned StoreSize = (unsigned)SizeInBits >> 3; const SCEVConstant *Stride = dyn_cast(StoreEv->getOperand(1)); if (Stride == 0 || StoreSize != Stride->getValue()->getValue()) { // TODO: Could also handle negative stride here someday, that will require // the validity check in mayLoopAccessLocation to be updated though. // Enable this to print exact negative strides. if (0 && Stride && StoreSize == -Stride->getValue()->getValue()) { dbgs() << "NEGATIVE STRIDE: " << *SI << "\n"; dbgs() << "BB: " << *SI->getParent(); } return false; } // See if we can optimize just this store in isolation. if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(), StoredVal, SI, StoreEv, BECount)) return true; // If the stored value is a strided load in the same loop with the same stride // this this may be transformable into a memcpy. This kicks in for stuff like // for (i) A[i] = B[i]; if (LoadInst *LI = dyn_cast(StoredVal)) { const SCEVAddRecExpr *LoadEv = dyn_cast(SE->getSCEV(LI->getOperand(0))); if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() && StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple()) if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount)) return true; } //errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n"; return false; } /// 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 with a constant size. if (MSI->isVolatile() || !isa(MSI->getLength())) return false; // If we're not allowed to hack on memset, we fail. if (!TLI->has(LibFunc::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 == 0 || Ev->getLoop() != CurLoop || !Ev->isAffine()) return false; // Reject memsets that are so large that they overflow an unsigned. uint64_t SizeInBytes = cast(MSI->getLength())->getZExtValue(); if ((SizeInBytes >> 32) != 0) return false; // Check to see if the stride matches the size of the memset. If so, then we // know that every byte is touched in the loop. const SCEVConstant *Stride = dyn_cast(Ev->getOperand(1)); // TODO: Could also handle negative stride here someday, that will require the // validity check in mayLoopAccessLocation to be updated though. if (Stride == 0 || MSI->getLength() != Stride->getValue()) return false; return processLoopStridedStore(Pointer, (unsigned)SizeInBytes, MSI->getAlignment(), MSI->getValue(), MSI, Ev, BECount); } /// 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,AliasAnalysis::ModRefResult Access, Loop *L, const SCEV *BECount, unsigned StoreSize, AliasAnalysis &AA, Instruction *IgnoredStore) { // 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. uint64_t AccessSize = AliasAnalysis::UnknownSize; // 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 if (const SCEVConstant *BECst = dyn_cast(BECount)) AccessSize = (BECst->getValue()->getZExtValue()+1)*StoreSize; // 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]. AliasAnalysis::Location StoreLoc(Ptr, AccessSize); for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E; ++BI) for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I) if (&*I != IgnoredStore && (AA.getModRefInfo(I, StoreLoc) & Access)) return true; return false; } /// 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 TargetData &TD) { // 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 == 0) return 0; // Only handle simple values that are a power of two bytes in size. uint64_t Size = TD.getTypeSizeInBits(V->getType()); if (Size == 0 || (Size & 7) || (Size & (Size-1))) return 0; // Don't care enough about darwin/ppc to implement this. if (TD.isBigEndian()) return 0; // 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 0; // 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)); } /// 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, unsigned StoreSize, unsigned StoreAlignment, Value *StoredVal, Instruction *TheStore, const SCEVAddRecExpr *Ev, const SCEV *BECount) { // 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); Constant *PatternValue = 0; // If we're allowed to form a memset, and the stored value would be acceptable // for memset, use it. if (SplatValue && TLI->has(LibFunc::memset) && // Verify that the stored value is loop invariant. If not, we can't // promote the memset. CurLoop->isLoopInvariant(SplatValue)) { // Keep and use SplatValue. PatternValue = 0; } else if (TLI->has(LibFunc::memset_pattern16) && (PatternValue = getMemSetPatternValue(StoredVal, *TD))) { // It looks like we can use PatternValue! SplatValue = 0; } else { // Otherwise, this isn't an idiom we can transform. For example, we can't // do anything with a 3-byte store. 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, "loop-idiom"); // 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. unsigned AddrSpace = cast(DestPtr->getType())->getAddressSpace(); Value *BasePtr = Expander.expandCodeFor(Ev->getStart(), Builder.getInt8PtrTy(AddrSpace), Preheader->getTerminator()); if (mayLoopAccessLocation(BasePtr, AliasAnalysis::ModRef, CurLoop, BECount, StoreSize, getAnalysis(), TheStore)){ Expander.clear(); // If we generated new code for the base pointer, clean up. deleteIfDeadInstruction(BasePtr, *SE); return false; } // Okay, everything looks good, insert the memset. // The # stored bytes is (BECount+1)*Size. Expand the trip count out to // pointer size if it isn't already. Type *IntPtr = TD->getIntPtrType(DestPtr->getContext()); BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr); const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1), SCEV::FlagNUW); if (StoreSize != 1) NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize), SCEV::FlagNUW); Value *NumBytes = Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator()); CallInst *NewCall; if (SplatValue) NewCall = Builder.CreateMemSet(BasePtr, SplatValue,NumBytes,StoreAlignment); else { Module *M = TheStore->getParent()->getParent()->getParent(); Value *MSP = M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(), Builder.getInt8PtrTy(), Builder.getInt8PtrTy(), IntPtr, (void*)0); // 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::InternalLinkage, PatternValue, ".memset_pattern"); GV->setUnnamedAddr(true); // Ok to merge these. GV->setAlignment(16); Value *PatternPtr = ConstantExpr::getBitCast(GV, Builder.getInt8PtrTy()); NewCall = Builder.CreateCall3(MSP, BasePtr, PatternPtr, NumBytes); } DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n" << " from store to: " << *Ev << " at: " << *TheStore << "\n"); NewCall->setDebugLoc(TheStore->getDebugLoc()); // Okay, the memset has been formed. Zap the original store and anything that // feeds into it. deleteDeadInstruction(TheStore, *SE); ++NumMemSet; return true; } /// processLoopStoreOfLoopLoad - We see a strided store whose value is a /// same-strided load. bool LoopIdiomRecognize:: processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize, const SCEVAddRecExpr *StoreEv, const SCEVAddRecExpr *LoadEv, const SCEV *BECount) { // If we're not allowed to form memcpy, we fail. if (!TLI->has(LibFunc::memcpy)) return false; LoadInst *LI = cast(SI->getValueOperand()); // 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, "loop-idiom"); // 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(StoreEv->getStart(), Builder.getInt8PtrTy(SI->getPointerAddressSpace()), Preheader->getTerminator()); if (mayLoopAccessLocation(StoreBasePtr, AliasAnalysis::ModRef, CurLoop, BECount, StoreSize, getAnalysis(), SI)) { Expander.clear(); // If we generated new code for the base pointer, clean up. deleteIfDeadInstruction(StoreBasePtr, *SE); return false; } // For a memcpy, we have to make sure that the input array is not being // mutated by the loop. Value *LoadBasePtr = Expander.expandCodeFor(LoadEv->getStart(), Builder.getInt8PtrTy(LI->getPointerAddressSpace()), Preheader->getTerminator()); if (mayLoopAccessLocation(LoadBasePtr, AliasAnalysis::Mod, CurLoop, BECount, StoreSize, getAnalysis(), SI)) { Expander.clear(); // If we generated new code for the base pointer, clean up. deleteIfDeadInstruction(LoadBasePtr, *SE); deleteIfDeadInstruction(StoreBasePtr, *SE); return false; } // Okay, everything is safe, we can transform this! // The # stored bytes is (BECount+1)*Size. Expand the trip count out to // pointer size if it isn't already. Type *IntPtr = TD->getIntPtrType(SI->getContext()); BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr); const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1), SCEV::FlagNUW); if (StoreSize != 1) NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize), SCEV::FlagNUW); Value *NumBytes = Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator()); CallInst *NewCall = Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes, std::min(SI->getAlignment(), LI->getAlignment())); NewCall->setDebugLoc(SI->getDebugLoc()); DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n" << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"); // Okay, the memset has been formed. Zap the original store and anything that // feeds into it. deleteDeadInstruction(SI, *SE); ++NumMemCpy; return true; }