X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTransforms%2FScalar%2FLoopIdiomRecognize.cpp;h=ee5203105201f791815302dafe72402785f30867;hp=b57b7c276d72511ec4e75daee415cc1459f50eeb;hb=62b7c9cb848c8f48a40d2697015bdaba0c789fe7;hpb=383c55870eda2ce1664d8e9d199b1ba7bacd32ba diff --git a/lib/Transforms/Scalar/LoopIdiomRecognize.cpp b/lib/Transforms/Scalar/LoopIdiomRecognize.cpp index b57b7c276d7..ee520310520 100644 --- a/lib/Transforms/Scalar/LoopIdiomRecognize.cpp +++ b/lib/Transforms/Scalar/LoopIdiomRecognize.cpp @@ -67,149 +67,77 @@ STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores"); namespace { - class LoopIdiomRecognize; - - /// This class defines some utility functions for loop idiom recognization. - class LIRUtil { - public: - /// Return true iff the block contains nothing but an uncondition branch - /// (aka goto instruction). - static bool isAlmostEmpty(BasicBlock *); +class LoopIdiomRecognize : public LoopPass { + Loop *CurLoop; + AliasAnalysis *AA; + DominatorTree *DT; + LoopInfo *LI; + ScalarEvolution *SE; + TargetLibraryInfo *TLI; + const TargetTransformInfo *TTI; + +public: + static char ID; + explicit LoopIdiomRecognize() : LoopPass(ID) { + initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry()); + } - static BranchInst *getBranch(BasicBlock *BB) { - return dyn_cast(BB->getTerminator()); - } + bool runOnLoop(Loop *L, LPPassManager &LPM) override; + + /// This transformation requires natural loop information & requires that + /// loop preheaders be inserted into the CFG. + /// + void getAnalysisUsage(AnalysisUsage &AU) const override { + 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(); + AU.addRequired(); + } - /// Derive the precondition block (i.e the block that guards the loop - /// preheader) from the given preheader. - static BasicBlock *getPrecondBb(BasicBlock *PreHead); - }; - - /// This class is to recoginize idioms of population-count conducted in - /// a noncountable loop. Currently it only recognizes this pattern: - /// \code - /// while(x) {cnt++; ...; x &= x - 1; ...} - /// \endcode - class NclPopcountRecognize { - LoopIdiomRecognize &LIR; - Loop *CurLoop; - BasicBlock *PreCondBB; - - typedef IRBuilder<> IRBuilderTy; - - public: - explicit NclPopcountRecognize(LoopIdiomRecognize &TheLIR); - bool recognize(); - - private: - /// Take a glimpse of the loop to see if we need to go ahead recoginizing - /// the idiom. - bool preliminaryScreen(); - - /// Check if the given conditional branch is based on the comparison - /// between a variable and zero, and if the variable is non-zero, the - /// control yields to the loop entry. If the branch matches the behavior, - /// the variable involved in the comparion is returned. This function will - /// be called to see if the precondition and postcondition of the loop - /// are in desirable form. - Value *matchCondition(BranchInst *Br, BasicBlock *NonZeroTarget) const; - - /// Return true iff the idiom is detected in the loop. and 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. - bool detectIdiom - (Instruction *&CntInst, PHINode *&CntPhi, Value *&Var) const; - - /// Insert ctpop intrinsic function and some obviously dead instructions. - void transform(Instruction *CntInst, PHINode *CntPhi, Value *Var); - - /// Create llvm.ctpop.* intrinsic function. - CallInst *createPopcntIntrinsic(IRBuilderTy &IRB, Value *Val, DebugLoc DL); - }; - - class LoopIdiomRecognize : public LoopPass { - Loop *CurLoop; - DominatorTree *DT; - ScalarEvolution *SE; - TargetLibraryInfo *TLI; - const TargetTransformInfo *TTI; - public: - static char ID; - explicit LoopIdiomRecognize() : LoopPass(ID) { - initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry()); - DT = nullptr; - SE = nullptr; - TLI = nullptr; - TTI = nullptr; - } +private: + /// \name Countable Loop Idiom Handling + /// @{ - bool runOnLoop(Loop *L, LPPassManager &LPM) override; - 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. - /// - void getAnalysisUsage(AnalysisUsage &AU) const override { - 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(); - AU.addRequired(); - } + bool runOnCountableLoop(); + bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount, + SmallVectorImpl &ExitBlocks); - DominatorTree *getDominatorTree() { - return DT ? DT - : (DT = &getAnalysis().getDomTree()); - } + bool processLoopStore(StoreInst *SI, const SCEV *BECount); + bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount); - ScalarEvolution *getScalarEvolution() { - return SE ? SE : (SE = &getAnalysis()); - } + 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); - TargetLibraryInfo *getTargetLibraryInfo() { - if (!TLI) - TLI = &getAnalysis().getTLI(); + /// @} + /// \name Noncountable Loop Idiom Handling + /// @{ - return TLI; - } + bool runOnNoncountableLoop(); - const TargetTransformInfo *getTargetTransformInfo() { - return TTI ? TTI - : (TTI = &getAnalysis().getTTI( - *CurLoop->getHeader()->getParent())); - } + bool recognizePopcount(); + void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst, + PHINode *CntPhi, Value *Var); - Loop *getLoop() const { return CurLoop; } + /// @} +}; - private: - bool runOnNoncountableLoop(); - bool runOnCountableLoop(); - }; -} +} // End anonymous namespace. char LoopIdiomRecognize::ID = 0; INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms", @@ -242,867 +170,823 @@ static void deleteDeadInstruction(Instruction *I, //===----------------------------------------------------------------------===// // -// Implementation of LIRUtil -// -//===----------------------------------------------------------------------===// - -// This function will return true iff the given block contains nothing but goto. -// A typical usage of this function is to check if the preheader function is -// "almost" empty such that generated intrinsic functions can be moved across -// the preheader and be placed at the end of the precondition block without -// the concern of breaking data dependence. -bool LIRUtil::isAlmostEmpty(BasicBlock *BB) { - if (BranchInst *Br = getBranch(BB)) { - return Br->isUnconditional() && Br == BB->begin(); - } - return false; -} - -BasicBlock *LIRUtil::getPrecondBb(BasicBlock *PreHead) { - if (BasicBlock *BB = PreHead->getSinglePredecessor()) { - BranchInst *Br = getBranch(BB); - return Br && Br->isConditional() ? BB : nullptr; - } - return nullptr; -} - -//===----------------------------------------------------------------------===// -// -// Implementation of NclPopcountRecognize +// Implementation of LoopIdiomRecognize // //===----------------------------------------------------------------------===// -NclPopcountRecognize::NclPopcountRecognize(LoopIdiomRecognize &TheLIR): - LIR(TheLIR), CurLoop(TheLIR.getLoop()), PreCondBB(nullptr) { -} - -bool NclPopcountRecognize::preliminaryScreen() { - const TargetTransformInfo *TTI = LIR.getTargetTransformInfo(); - if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware) +bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) { + if (skipOptnoneFunction(L)) return false; - // Counting population are usually conducted by few arithmetic instructions. - // Such instructions can be easilly "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) + 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; - BasicBlock *LoopBody = *(CurLoop->block_begin()); - if (LoopBody->size() >= 20) { - // The loop is too big, bail out. + // 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; - } - // It should have a preheader containing nothing but a goto instruction. - BasicBlock *PreHead = CurLoop->getLoopPreheader(); - if (!PreHead || !LIRUtil::isAlmostEmpty(PreHead)) - return false; + AA = &getAnalysis(); + DT = &getAnalysis().getDomTree(); + LI = &getAnalysis().getLoopInfo(); + SE = &getAnalysis(); + TLI = &getAnalysis().getTLI(); + TTI = &getAnalysis().getTTI( + *CurLoop->getHeader()->getParent()); - // It should have a precondition block where the generated popcount instrinsic - // function will be inserted. - PreCondBB = LIRUtil::getPrecondBb(PreHead); - if (!PreCondBB) - return false; + if (SE->hasLoopInvariantBackedgeTakenCount(L)) + return runOnCountableLoop(); - return true; + return runOnNoncountableLoop(); } -Value *NclPopcountRecognize::matchCondition(BranchInst *Br, - BasicBlock *LoopEntry) const { - if (!Br || !Br->isConditional()) - return nullptr; - - ICmpInst *Cond = dyn_cast(Br->getCondition()); - if (!Cond) - return nullptr; - - ConstantInt *CmpZero = dyn_cast(Cond->getOperand(1)); - if (!CmpZero || !CmpZero->isZero()) - return nullptr; - - ICmpInst::Predicate Pred = Cond->getPredicate(); - if ((Pred == ICmpInst::ICMP_NE && Br->getSuccessor(0) == LoopEntry) || - (Pred == ICmpInst::ICMP_EQ && Br->getSuccessor(1) == LoopEntry)) - return Cond->getOperand(0); +bool LoopIdiomRecognize::runOnCountableLoop() { + const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop); + assert(!isa(BECount) && + "runOnCountableLoop() called on a loop without a predictable" + "backedge-taken count"); - return nullptr; -} + // 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; -bool NclPopcountRecognize::detectIdiom(Instruction *&CntInst, - PHINode *&CntPhi, - Value *&Var) const { - // Following code tries to detect this idiom: - // - // 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: - // + SmallVector ExitBlocks; + CurLoop->getUniqueExitBlocks(ExitBlocks); - // 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; + DEBUG(dbgs() << "loop-idiom Scanning: F[" + << CurLoop->getHeader()->getParent()->getName() << "] Loop %" + << CurLoop->getHeader()->getName() << "\n"); - DefX2 = CountInst = nullptr; - VarX1 = VarX0 = nullptr; - PhiX = CountPhi = nullptr; - LoopEntry = *(CurLoop->block_begin()); + 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; - // step 1: Check if the loop-back branch is in desirable form. - { - if (Value *T = matchCondition (LIRUtil::getBranch(LoopEntry), LoopEntry)) - DefX2 = dyn_cast(T); - else - return false; + MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks); } + return MadeChange; +} - // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)" - { - if (!DefX2 || DefX2->getOpcode() != Instruction::And) +/// 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; - BinaryOperator *SubOneOp; + 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 ((SubOneOp = dyn_cast(DefX2->getOperand(0)))) - VarX1 = DefX2->getOperand(1); - else { - VarX1 = DefX2->getOperand(0); - SubOneOp = dyn_cast(DefX2->getOperand(1)); + // If processing the store invalidated our iterator, start over from the + // top of the block. + if (!InstPtr) + I = BB->begin(); + continue; } - if (!SubOneOp) - return false; - Instruction *SubInst = cast(SubOneOp); - ConstantInt *Dec = dyn_cast(SubInst->getOperand(1)); - if (!Dec || - !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) || - (SubInst->getOpcode() == Instruction::Add && Dec->isAllOnesValue()))) { - return false; - } - } + // 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; - // step 3: Check the recurrence of variable X - { - PhiX = dyn_cast(VarX1); - if (!PhiX || - (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) { - return false; + // If processing the memset invalidated our iterator, start over from the + // top of the block. + if (!InstPtr) + I = BB->begin(); + continue; } } - // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1 - { - CountInst = nullptr; - for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI(), - IterE = LoopEntry->end(); Iter != IterE; Iter++) { - Instruction *Inst = Iter; - if (Inst->getOpcode() != Instruction::Add) - continue; + return MadeChange; +} - ConstantInt *Inc = dyn_cast(Inst->getOperand(1)); - if (!Inc || !Inc->isOne()) - continue; +/// 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; - PHINode *Phi = dyn_cast(Inst->getOperand(0)); - if (!Phi || Phi->getParent() != LoopEntry) - continue; + Value *StoredVal = SI->getValueOperand(); + Value *StorePtr = SI->getPointerOperand(); - // 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; - } - } + // Reject stores that are so large that they overflow an unsigned. + auto &DL = CurLoop->getHeader()->getModule()->getDataLayout(); + uint64_t SizeInBits = DL.getTypeSizeInBits(StoredVal->getType()); + if ((SizeInBits & 7) || (SizeInBits >> 32) != 0) + return false; - if (LiveOutLoop) { - CountInst = Inst; - CountPhi = Phi; - break; - } + // 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 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 || 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(); } - if (!CountInst) - return false; + 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;" - { - BranchInst *PreCondBr = LIRUtil::getBranch(PreCondBB); - Value *T = matchCondition (PreCondBr, CurLoop->getLoopPreheader()); - if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1)) - return false; + // See if we can optimize just this store in isolation. + if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(), + StoredVal, SI, StoreEv, BECount)) + return true; - CntInst = CountInst; - CntPhi = CountPhi; - Var = T; + // 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 true; + return false; } -void NclPopcountRecognize::transform(Instruction *CntInst, - PHINode *CntPhi, Value *Var) { +/// 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; - ScalarEvolution *SE = LIR.getScalarEvolution(); - TargetLibraryInfo *TLI = LIR.getTargetLibraryInfo(); - BasicBlock *PreHead = CurLoop->getLoopPreheader(); - BranchInst *PreCondBr = LIRUtil::getBranch(PreCondBB); - const DebugLoc DL = CntInst->getDebugLoc(); + // If we're not allowed to hack on memset, we fail. + if (!TLI->has(LibFunc::memset)) + return false; - // Assuming before transformation, the loop is following: - // if (x) // the precondition - // do { cnt++; x &= x - 1; } while(x); + Value *Pointer = MSI->getDest(); - // Step 1: Insert the ctpop instruction at the end of the precondition block - IRBuilderTy 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); + // 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 || !Ev->isAffine()) + return false; - // TripCnt is exactly the number of iterations the loop has - TripCnt = NewCount; + // Reject memsets that are so large that they overflow an unsigned. + uint64_t SizeInBytes = cast(MSI->getLength())->getZExtValue(); + if ((SizeInBytes >> 32) != 0) + return false; - // 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); - } - } + // 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)); - // Step 2: Replace the precondition from "if(x == 0) goto loop-exit" to - // "if(NewCount == 0) loop-exit". Withtout 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()); + // TODO: Could also handle negative stride here someday, that will require the + // validity check in mayLoopAccessLocation to be updated though. + if (!Stride || MSI->getLength() != Stride->getValue()) + return false; - Value *Opnd0 = PopCntZext; - Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0); - if (PreCond->getOperand(0) != Var) - std::swap(Opnd0, Opnd1); + return processLoopStridedStore(Pointer, (unsigned)SizeInBytes, + MSI->getAlignment(), MSI->getValue(), MSI, Ev, + BECount); +} - ICmpInst *NewPreCond = - cast(Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1)); - PreCond->replaceAllUsesWith(NewPreCond); +/// 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, 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 = MemoryLocation::UnknownSize; - RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI); - } + // 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; - // Step 3: Note that the population count is exactly the trip count of the - // loop in question, which enble us to to convert the loop from noncountable - // loop into a countable one. The benefit is twofold: - // - // - If the loop only counts population, the entire loop become dead after - // the transformation. It is lots easier to prove a countable loop dead - // than to prove a noncountable one. (In some C dialects, a infite 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()); - { - BranchInst *LbBr = LIRUtil::getBranch(Body); - ICmpInst *LbCond = cast(LbBr->getCondition()); - Type *Ty = TripCnt->getType(); + // 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); - PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", Body->begin()); + 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; - Builder.SetInsertPoint(LbCond); - Value *Opnd1 = cast(TcPhi); - Value *Opnd2 = cast(ConstantInt::get(Ty, 1)); - Instruction *TcDec = - cast(Builder.CreateSub(Opnd1, Opnd2, "tcdec", false, true)); + return false; +} - TcPhi->addIncoming(TripCnt, PreHead); - TcPhi->addIncoming(TcDec, Body); +/// 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) { + // 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) + return nullptr; - CmpInst::Predicate Pred = (LbBr->getSuccessor(0) == Body) ? - CmpInst::ICMP_UGT : CmpInst::ICMP_SLE; - LbCond->setPredicate(Pred); - LbCond->setOperand(0, TcDec); - LbCond->setOperand(1, cast(ConstantInt::get(Ty, 0))); - } + // 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; - // 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); + // Don't care enough about darwin/ppc to implement this. + if (DL.isBigEndian()) + return nullptr; - // 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); -} + // Convert to size in bytes. + Size /= 8; -CallInst *NclPopcountRecognize::createPopcntIntrinsic(IRBuilderTy &IRBuilder, - Value *Val, DebugLoc DL) { - Value *Ops[] = { Val }; - Type *Tys[] = { Val->getType() }; + // 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; - Module *M = (*(CurLoop->block_begin()))->getParent()->getParent(); - Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys); - CallInst *CI = IRBuilder.CreateCall(Func, Ops); - CI->setDebugLoc(DL); + // If the constant is exactly 16 bytes, just use it. + if (Size == 16) + return C; - return CI; + // 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)); } -/// recognize - detect population count idiom in a non-countable loop. If -/// detected, transform the relevant code to popcount intrinsic function -/// call, and return true; otherwise, return false. -bool NclPopcountRecognize::recognize() { - - if (!LIR.getTargetTransformInfo()) - return false; - - LIR.getScalarEvolution(); +/// 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 (!preliminaryScreen()) - return false; + // 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 = nullptr; + auto &DL = CurLoop->getHeader()->getModule()->getDataLayout(); + unsigned DestAS = DestPtr->getType()->getPointerAddressSpace(); - Instruction *CntInst; - PHINode *CntPhi; - Value *Val; - if (!detectIdiom(CntInst, CntPhi, Val)) + // 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 = nullptr; + } else if (DestAS == 0 && TLI->has(LibFunc::memset_pattern16) && + (PatternValue = getMemSetPatternValue(StoredVal, DL))) { + // Don't create memset_pattern16s with address spaces. + // It looks like we can use PatternValue! + SplatValue = nullptr; + } else { + // Otherwise, this isn't an idiom we can transform. For example, we can't + // do anything with a 3-byte store. return false; + } - transform(CntInst, CntPhi, Val); - return true; -} + // 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"); -//===----------------------------------------------------------------------===// -// -// Implementation of LoopIdiomRecognize -// -//===----------------------------------------------------------------------===// + Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS); -bool LoopIdiomRecognize::runOnCountableLoop() { - const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop); - assert(!isa(BECount) && - "runOnCountableLoop() called on a loop without a predictable" - "backedge-taken count"); + // 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(Ev->getStart(), DestInt8PtrTy, + Preheader->getTerminator()); - // 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; + if (mayLoopAccessLocation(BasePtr, MRI_ModRef, CurLoop, BECount, StoreSize, + *AA, TheStore)) { + Expander.clear(); + // If we generated new code for the base pointer, clean up. + RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI); + return false; + } - // set DT - (void)getDominatorTree(); + // Okay, everything looks good, insert the memset. - LoopInfo &LI = getAnalysis().getLoopInfo(); - TLI = &getAnalysis().getTLI(); + // The # stored bytes is (BECount+1)*Size. Expand the trip count out to + // pointer size if it isn't already. + Type *IntPtr = Builder.getIntPtrTy(DL, DestAS); + BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr); - // set TLI - (void)getTargetLibraryInfo(); + 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); + } - SmallVector ExitBlocks; - CurLoop->getUniqueExitBlocks(ExitBlocks); + Value *NumBytes = + Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator()); - DEBUG(dbgs() << "loop-idiom Scanning: F[" - << CurLoop->getHeader()->getParent()->getName() - << "] Loop %" << CurLoop->getHeader()->getName() << "\n"); + CallInst *NewCall; + if (SplatValue) { + NewCall = + Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment); + } else { + // Everything is emitted in default address space + Type *Int8PtrTy = DestInt8PtrTy; - bool MadeChange = false; - // Scan all the blocks in the loop that are not in subloops. - for (Loop::block_iterator BI = CurLoop->block_begin(), - E = CurLoop->block_end(); BI != E; ++BI) { - // Ignore blocks in subloops. - if (LI.getLoopFor(*BI) != CurLoop) - continue; + Module *M = TheStore->getParent()->getParent()->getParent(); + Value *MSP = + M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(), + Int8PtrTy, Int8PtrTy, IntPtr, (void *)nullptr); - MadeChange |= runOnLoopBlock(*BI, BECount, ExitBlocks); + // 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(true); // Ok to merge these. + GV->setAlignment(16); + Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy); + NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes}); } - return MadeChange; -} -bool LoopIdiomRecognize::runOnNoncountableLoop() { - NclPopcountRecognize Popcount(*this); - if (Popcount.recognize()) - return true; + DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n" + << " from store to: " << *Ev << " at: " << *TheStore << "\n"); + NewCall->setDebugLoc(TheStore->getDebugLoc()); - return false; + // Okay, the memset has been formed. Zap the original store and anything that + // feeds into it. + deleteDeadInstruction(TheStore, TLI); + ++NumMemSet; + return true; } -bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) { - if (skipOptnoneFunction(L)) +/// 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; - CurLoop = L; + LoadInst *LI = cast(SI->getValueOperand()); - // 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; - - SE = &getAnalysis(); - if (SE->hasLoopInvariantBackedgeTakenCount(L)) - return runOnCountableLoop(); - return runOnNoncountableLoop(); -} - -/// 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) - I = BB->begin(); - continue; - } + // 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()); + const DataLayout &DL = Preheader->getModule()->getDataLayout(); + SCEVExpander Expander(*SE, DL, "loop-idiom"); - // 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; + // 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 processing the memset invalidated our iterator, start over from the - // top of the block. - if (!InstPtr) - I = BB->begin(); - continue; - } + if (mayLoopAccessLocation(StoreBasePtr, MRI_ModRef, CurLoop, BECount, + StoreSize, *AA, SI)) { + Expander.clear(); + // If we generated new code for the base pointer, clean up. + RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI); + return false; } - return MadeChange; -} + // 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, MRI_Mod, CurLoop, BECount, StoreSize, + *AA, SI)) { + Expander.clear(); + // If we generated new code for the base pointer, clean up. + RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI); + RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI); + return false; + } -/// 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; + // Okay, everything is safe, we can transform this! - Value *StoredVal = SI->getValueOperand(); - Value *StorePtr = SI->getPointerOperand(); + // The # stored bytes is (BECount+1)*Size. Expand the trip count out to + // pointer size if it isn't already. + Type *IntPtrTy = Builder.getIntPtrTy(DL, SI->getPointerAddressSpace()); + BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy); - // Reject stores that are so large that they overflow an unsigned. - auto &DL = CurLoop->getHeader()->getModule()->getDataLayout(); - uint64_t SizeInBits = DL.getTypeSizeInBits(StoredVal->getType()); - if ((SizeInBits & 7) || (SizeInBits >> 32) != 0) - return false; + const SCEV *NumBytesS = + SE->getAddExpr(BECount, SE->getConstant(IntPtrTy, 1), SCEV::FlagNUW); + if (StoreSize != 1) + NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize), + SCEV::FlagNUW); - // 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 false; + Value *NumBytes = + Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator()); - // 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)); + CallInst *NewCall = + Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes, + std::min(SI->getAlignment(), LI->getAlignment())); + NewCall->setDebugLoc(SI->getDebugLoc()); - if (!Stride || 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(); - } + DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n" + << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" + << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"); - return false; - } + // Okay, the memset has been formed. Zap the original store and anything that + // feeds into it. + deleteDeadInstruction(SI, TLI); + ++NumMemCpy; + return true; +} - // See if we can optimize just this store in isolation. - if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(), - StoredVal, SI, StoreEv, BECount)) +bool LoopIdiomRecognize::runOnNoncountableLoop() { + if (recognizePopcount()) 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 || Ev->getLoop() != CurLoop || !Ev->isAffine()) - return false; +/// Check if the given conditional branch is based on the comparison between +/// a variable and zero, and if the variable is non-zero, the control yields to +/// the loop entry. If the branch matches the behavior, the variable involved +/// in the comparion 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) { + if (!BI || !BI->isConditional()) + return nullptr; - // Reject memsets that are so large that they overflow an unsigned. - uint64_t SizeInBytes = cast(MSI->getLength())->getZExtValue(); - if ((SizeInBytes >> 32) != 0) - return false; + ICmpInst *Cond = dyn_cast(BI->getCondition()); + if (!Cond) + return nullptr; - // 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)); + ConstantInt *CmpZero = dyn_cast(Cond->getOperand(1)); + if (!CmpZero || !CmpZero->isZero()) + return nullptr; - // TODO: Could also handle negative stride here someday, that will require the - // validity check in mayLoopAccessLocation to be updated though. - if (!Stride || MSI->getLength() != Stride->getValue()) - return false; + ICmpInst::Predicate Pred = Cond->getPredicate(); + if ((Pred == ICmpInst::ICMP_NE && BI->getSuccessor(0) == LoopEntry) || + (Pred == ICmpInst::ICMP_EQ && BI->getSuccessor(1) == LoopEntry)) + return Cond->getOperand(0); - return processLoopStridedStore(Pointer, (unsigned)SizeInBytes, - MSI->getAlignment(), MSI->getValue(), - MSI, Ev, BECount); + 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; -/// 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; + DefX2 = CountInst = nullptr; + VarX1 = VarX0 = nullptr; + PhiX = CountPhi = nullptr; + LoopEntry = *(CurLoop->block_begin()); - // 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; + // 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; + } - // 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); + // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)" + { + if (!DefX2 || DefX2->getOpcode() != Instruction::And) + return false; - 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; + BinaryOperator *SubOneOp; - return false; -} + if ((SubOneOp = dyn_cast(DefX2->getOperand(0)))) + VarX1 = DefX2->getOperand(1); + else { + VarX1 = DefX2->getOperand(0); + SubOneOp = dyn_cast(DefX2->getOperand(1)); + } + if (!SubOneOp) + 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 DataLayout &DL) { - // 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) return nullptr; + Instruction *SubInst = cast(SubOneOp); + ConstantInt *Dec = dyn_cast(SubInst->getOperand(1)); + if (!Dec || + !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) || + (SubInst->getOpcode() == Instruction::Add && + Dec->isAllOnesValue()))) { + return false; + } + } - // 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; + // step 3: Check the recurrence of variable X + { + PhiX = dyn_cast(VarX1); + if (!PhiX || + (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) { + return false; + } + } - // Don't care enough about darwin/ppc to implement this. - if (DL.isBigEndian()) - return nullptr; + // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1 + { + CountInst = nullptr; + for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI(), + IterE = LoopEntry->end(); + Iter != IterE; Iter++) { + Instruction *Inst = Iter; + if (Inst->getOpcode() != Instruction::Add) + continue; - // Convert to size in bytes. - Size /= 8; + ConstantInt *Inc = dyn_cast(Inst->getOperand(1)); + if (!Inc || !Inc->isOne()) + continue; - // 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; + PHINode *Phi = dyn_cast(Inst->getOperand(0)); + if (!Phi || Phi->getParent() != LoopEntry) + 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 the constant is exactly 16 bytes, just use it. - if (Size == 16) return C; + if (LiveOutLoop) { + CountInst = Inst; + CountPhi = Phi; + break; + } + } - // 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)); -} + 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; -/// 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) { + CntInst = CountInst; + CntPhi = CountPhi; + Var = T; + } - // 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 = nullptr; - auto &DL = CurLoop->getHeader()->getModule()->getDataLayout(); - unsigned DestAS = DestPtr->getType()->getPointerAddressSpace(); + return true; +} - // 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 = nullptr; - } else if (DestAS == 0 && TLI->has(LibFunc::memset_pattern16) && - (PatternValue = getMemSetPatternValue(StoredVal, DL))) { - // Don't create memset_pattern16s with address spaces. - // It looks like we can use PatternValue! - SplatValue = nullptr; - } else { - // Otherwise, this isn't an idiom we can transform. For example, we can't - // do anything with a 3-byte store. +/// 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; - } - - // 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"); - Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS); + // Counting population are usually conducted by few arithmetic instructions. + // Such instructions can be easilly "absorbed" by vacant slots in a + // non-compact loop. Therefore, recognizing popcount idiom only makes sense + // in a compact loop. - // 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(Ev->getStart(), DestInt8PtrTy, - Preheader->getTerminator()); + // Give up if the loop has multiple blocks or multiple backedges. + if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1) + return false; - if (mayLoopAccessLocation(BasePtr, AliasAnalysis::ModRef, - CurLoop, BECount, - StoreSize, getAnalysis(), TheStore)) { - Expander.clear(); - // If we generated new code for the base pointer, clean up. - RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI); + BasicBlock *LoopBody = *(CurLoop->block_begin()); + if (LoopBody->size() >= 20) { + // The loop is too big, bail out. return false; } - // Okay, everything looks good, insert the memset. + // It should have a preheader containing nothing but an unconditional branch. + BasicBlock *PH = CurLoop->getLoopPreheader(); + if (!PH) + return false; + if (&PH->front() != PH->getTerminator()) + return false; + auto *EntryBI = dyn_cast(PH->getTerminator()); + if (!EntryBI || EntryBI->isConditional()) + return false; - // The # stored bytes is (BECount+1)*Size. Expand the trip count out to - // pointer size if it isn't already. - Type *IntPtr = Builder.getIntPtrTy(DL, DestAS); - BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr); + // It should have a precondition block where the generated popcount instrinsic + // function can be inserted. + auto *PreCondBB = PH->getSinglePredecessor(); + if (!PreCondBB) + return false; + auto *PreCondBI = dyn_cast(PreCondBB->getTerminator()); + if (!PreCondBI || PreCondBI->isUnconditional()) + return false; - 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); - } + Instruction *CntInst; + PHINode *CntPhi; + Value *Val; + if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val)) + return false; - Value *NumBytes = - Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator()); + transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val); + return true; +} - CallInst *NewCall; - if (SplatValue) { - NewCall = Builder.CreateMemSet(BasePtr, - SplatValue, - NumBytes, - StoreAlignment); - } else { - // Everything is emitted in default address space - Type *Int8PtrTy = DestInt8PtrTy; +static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val, + DebugLoc DL) { + Value *Ops[] = {Val}; + Type *Tys[] = {Val->getType()}; - Module *M = TheStore->getParent()->getParent()->getParent(); - Value *MSP = M->getOrInsertFunction("memset_pattern16", - Builder.getVoidTy(), - Int8PtrTy, - Int8PtrTy, - IntPtr, - (void*)nullptr); + Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent(); + Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys); + CallInst *CI = IRBuilder.CreateCall(Func, Ops); + CI->setDebugLoc(DL); - // 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(true); // Ok to merge these. - GV->setAlignment(16); - Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy); - NewCall = Builder.CreateCall3(MSP, BasePtr, PatternPtr, NumBytes); - } + return CI; +} - DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n" - << " from store to: " << *Ev << " at: " << *TheStore << "\n"); - NewCall->setDebugLoc(TheStore->getDebugLoc()); +void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB, + Instruction *CntInst, + PHINode *CntPhi, Value *Var) { + BasicBlock *PreHead = CurLoop->getLoopPreheader(); + auto *PreCondBr = dyn_cast(PreCondBB->getTerminator()); + const DebugLoc DL = CntInst->getDebugLoc(); - // Okay, the memset has been formed. Zap the original store and anything that - // feeds into it. - deleteDeadInstruction(TheStore, TLI); - ++NumMemSet; - return true; -} + // Assuming before transformation, the loop is following: + // if (x) // the precondition + // do { cnt++; x &= x - 1; } while(x); -/// 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; + // 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())); - LoadInst *LI = cast(SI->getValueOperand()); + if (NewCount != PopCnt) + (cast(NewCount))->setDebugLoc(DL); - // 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()); - const DataLayout &DL = Preheader->getModule()->getDataLayout(); - SCEVExpander Expander(*SE, DL, "loop-idiom"); + // TripCnt is exactly the number of iterations the loop has + TripCnt = NewCount; - // 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. - RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI); - return false; + // 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); + } } - // 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()); + // Step 2: Replace the precondition from "if(x == 0) goto loop-exit" to + // "if(NewCount == 0) loop-exit". Withtout 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()); - if (mayLoopAccessLocation(LoadBasePtr, AliasAnalysis::Mod, CurLoop, BECount, - StoreSize, getAnalysis(), SI)) { - Expander.clear(); - // If we generated new code for the base pointer, clean up. - RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI); - RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI); - return false; - } + Value *Opnd0 = PopCntZext; + Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0); + if (PreCond->getOperand(0) != Var) + std::swap(Opnd0, Opnd1); - // Okay, everything is safe, we can transform this! + ICmpInst *NewPreCond = cast( + Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1)); + PreCondBr->setCondition(NewPreCond); + RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI); + } - // The # stored bytes is (BECount+1)*Size. Expand the trip count out to - // pointer size if it isn't already. - Type *IntPtrTy = Builder.getIntPtrTy(DL, SI->getPointerAddressSpace()); - BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy); + // Step 3: Note that the population count is exactly the trip count of the + // loop in question, which enble us to to convert the loop from noncountable + // loop into a countable one. The benefit is twofold: + // + // - If the loop only counts population, the entire loop become dead after + // the transformation. It is lots easier to prove a countable loop dead + // than to prove a noncountable one. (In some C dialects, a infite 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 = dyn_cast(Body->getTerminator()); + ICmpInst *LbCond = cast(LbBr->getCondition()); + Type *Ty = TripCnt->getType(); - const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtrTy, 1), - SCEV::FlagNUW); - if (StoreSize != 1) - NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize), - SCEV::FlagNUW); + PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", Body->begin()); - Value *NumBytes = - Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator()); + Builder.SetInsertPoint(LbCond); + Value *Opnd1 = cast(TcPhi); + Value *Opnd2 = cast(ConstantInt::get(Ty, 1)); + Instruction *TcDec = cast( + Builder.CreateSub(Opnd1, Opnd2, "tcdec", false, true)); - CallInst *NewCall = - Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes, - std::min(SI->getAlignment(), LI->getAlignment())); - NewCall->setDebugLoc(SI->getDebugLoc()); + TcPhi->addIncoming(TripCnt, PreHead); + TcPhi->addIncoming(TcDec, Body); - DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n" - << " from load ptr=" << *LoadEv << " at: " << *LI << "\n" - << " from store ptr=" << *StoreEv << " at: " << *SI << "\n"); + CmpInst::Predicate Pred = + (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE; + LbCond->setPredicate(Pred); + LbCond->setOperand(0, TcDec); + LbCond->setOperand(1, cast(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); - // Okay, the memset has been formed. Zap the original store and anything that - // feeds into it. - deleteDeadInstruction(SI, TLI); - ++NumMemCpy; - return true; + // 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); }