1 //===-- LoopIdiomRecognize.cpp - Loop idiom recognition -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass implements an idiom recognizer that transforms simple loops into a
11 // non-loop form. In cases that this kicks in, it can be a significant
14 //===----------------------------------------------------------------------===//
18 // Future loop memory idioms to recognize:
19 // memcmp, memmove, strlen, etc.
20 // Future floating point idioms to recognize in -ffast-math mode:
22 // Future integer operation idioms to recognize:
25 // Beware that isel's default lowering for ctpop is highly inefficient for
26 // i64 and larger types when i64 is legal and the value has few bits set. It
27 // would be good to enhance isel to emit a loop for ctpop in this case.
29 // We should enhance the memset/memcpy recognition to handle multiple stores in
30 // the loop. This would handle things like:
31 // void foo(_Complex float *P)
32 // for (i) { __real__(*P) = 0; __imag__(*P) = 0; }
34 // We should enhance this to handle negative strides through memory.
35 // Alternatively (and perhaps better) we could rely on an earlier pass to force
36 // forward iteration through memory, which is generally better for cache
37 // behavior. Negative strides *do* happen for memset/memcpy loops.
39 // This could recognize common matrix multiplies and dot product idioms and
40 // replace them with calls to BLAS (if linked in??).
42 //===----------------------------------------------------------------------===//
44 #define DEBUG_TYPE "loop-idiom"
45 #include "llvm/Transforms/Scalar.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/Analysis/AliasAnalysis.h"
48 #include "llvm/Analysis/LoopPass.h"
49 #include "llvm/Analysis/ScalarEvolutionExpander.h"
50 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
51 #include "llvm/Analysis/TargetTransformInfo.h"
52 #include "llvm/Analysis/ValueTracking.h"
53 #include "llvm/IR/DataLayout.h"
54 #include "llvm/IR/Dominators.h"
55 #include "llvm/IR/IRBuilder.h"
56 #include "llvm/IR/IntrinsicInst.h"
57 #include "llvm/IR/Module.h"
58 #include "llvm/Support/Debug.h"
59 #include "llvm/Support/raw_ostream.h"
60 #include "llvm/Target/TargetLibraryInfo.h"
61 #include "llvm/Transforms/Utils/Local.h"
64 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
65 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
69 class LoopIdiomRecognize;
71 /// This class defines some utility functions for loop idiom recognization.
74 /// Return true iff the block contains nothing but an uncondition branch
75 /// (aka goto instruction).
76 static bool isAlmostEmpty(BasicBlock *);
78 static BranchInst *getBranch(BasicBlock *BB) {
79 return dyn_cast<BranchInst>(BB->getTerminator());
82 /// Derive the precondition block (i.e the block that guards the loop
83 /// preheader) from the given preheader.
84 static BasicBlock *getPrecondBb(BasicBlock *PreHead);
87 /// This class is to recoginize idioms of population-count conducted in
88 /// a noncountable loop. Currently it only recognizes this pattern:
90 /// while(x) {cnt++; ...; x &= x - 1; ...}
92 class NclPopcountRecognize {
93 LoopIdiomRecognize &LIR;
95 BasicBlock *PreCondBB;
97 typedef IRBuilder<> IRBuilderTy;
100 explicit NclPopcountRecognize(LoopIdiomRecognize &TheLIR);
104 /// Take a glimpse of the loop to see if we need to go ahead recoginizing
106 bool preliminaryScreen();
108 /// Check if the given conditional branch is based on the comparison
109 /// between a variable and zero, and if the variable is non-zero, the
110 /// control yields to the loop entry. If the branch matches the behavior,
111 /// the variable involved in the comparion is returned. This function will
112 /// be called to see if the precondition and postcondition of the loop
113 /// are in desirable form.
114 Value *matchCondition (BranchInst *Br, BasicBlock *NonZeroTarget) const;
116 /// Return true iff the idiom is detected in the loop. and 1) \p CntInst
117 /// is set to the instruction counting the pupulation bit. 2) \p CntPhi
118 /// is set to the corresponding phi node. 3) \p Var is set to the value
119 /// whose population bits are being counted.
121 (Instruction *&CntInst, PHINode *&CntPhi, Value *&Var) const;
123 /// Insert ctpop intrinsic function and some obviously dead instructions.
124 void transform (Instruction *CntInst, PHINode *CntPhi, Value *Var);
126 /// Create llvm.ctpop.* intrinsic function.
127 CallInst *createPopcntIntrinsic(IRBuilderTy &IRB, Value *Val, DebugLoc DL);
130 class LoopIdiomRecognize : public LoopPass {
132 const DataLayout *DL;
135 TargetLibraryInfo *TLI;
136 const TargetTransformInfo *TTI;
139 explicit LoopIdiomRecognize() : LoopPass(ID) {
140 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
141 DL = 0; DT = 0; SE = 0; TLI = 0; TTI = 0;
144 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
145 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
146 SmallVectorImpl<BasicBlock*> &ExitBlocks);
148 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
149 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
151 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
152 unsigned StoreAlignment,
153 Value *SplatValue, Instruction *TheStore,
154 const SCEVAddRecExpr *Ev,
155 const SCEV *BECount);
156 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
157 const SCEVAddRecExpr *StoreEv,
158 const SCEVAddRecExpr *LoadEv,
159 const SCEV *BECount);
161 /// This transformation requires natural loop information & requires that
162 /// loop preheaders be inserted into the CFG.
164 void getAnalysisUsage(AnalysisUsage &AU) const override {
165 AU.addRequired<LoopInfo>();
166 AU.addPreserved<LoopInfo>();
167 AU.addRequiredID(LoopSimplifyID);
168 AU.addPreservedID(LoopSimplifyID);
169 AU.addRequiredID(LCSSAID);
170 AU.addPreservedID(LCSSAID);
171 AU.addRequired<AliasAnalysis>();
172 AU.addPreserved<AliasAnalysis>();
173 AU.addRequired<ScalarEvolution>();
174 AU.addPreserved<ScalarEvolution>();
175 AU.addPreserved<DominatorTreeWrapperPass>();
176 AU.addRequired<DominatorTreeWrapperPass>();
177 AU.addRequired<TargetLibraryInfo>();
178 AU.addRequired<TargetTransformInfo>();
181 const DataLayout *getDataLayout() {
184 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
185 DL = DLP ? &DLP->getDataLayout() : 0;
189 DominatorTree *getDominatorTree() {
191 : (DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree());
194 ScalarEvolution *getScalarEvolution() {
195 return SE ? SE : (SE = &getAnalysis<ScalarEvolution>());
198 TargetLibraryInfo *getTargetLibraryInfo() {
199 return TLI ? TLI : (TLI = &getAnalysis<TargetLibraryInfo>());
202 const TargetTransformInfo *getTargetTransformInfo() {
203 return TTI ? TTI : (TTI = &getAnalysis<TargetTransformInfo>());
206 Loop *getLoop() const { return CurLoop; }
209 bool runOnNoncountableLoop();
210 bool runOnCountableLoop();
214 char LoopIdiomRecognize::ID = 0;
215 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
217 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
218 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
219 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
220 INITIALIZE_PASS_DEPENDENCY(LCSSA)
221 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
222 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
223 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
224 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
225 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
228 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
230 /// deleteDeadInstruction - Delete this instruction. Before we do, go through
231 /// and zero out all the operands of this instruction. If any of them become
232 /// dead, delete them and the computation tree that feeds them.
234 static void deleteDeadInstruction(Instruction *I, ScalarEvolution &SE,
235 const TargetLibraryInfo *TLI) {
236 SmallVector<Instruction*, 32> NowDeadInsts;
238 NowDeadInsts.push_back(I);
240 // Before we touch this instruction, remove it from SE!
242 Instruction *DeadInst = NowDeadInsts.pop_back_val();
244 // This instruction is dead, zap it, in stages. Start by removing it from
246 SE.forgetValue(DeadInst);
248 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
249 Value *Op = DeadInst->getOperand(op);
250 DeadInst->setOperand(op, 0);
252 // If this operand just became dead, add it to the NowDeadInsts list.
253 if (!Op->use_empty()) continue;
255 if (Instruction *OpI = dyn_cast<Instruction>(Op))
256 if (isInstructionTriviallyDead(OpI, TLI))
257 NowDeadInsts.push_back(OpI);
260 DeadInst->eraseFromParent();
262 } while (!NowDeadInsts.empty());
265 /// deleteIfDeadInstruction - If the specified value is a dead instruction,
266 /// delete it and any recursively used instructions.
267 static void deleteIfDeadInstruction(Value *V, ScalarEvolution &SE,
268 const TargetLibraryInfo *TLI) {
269 if (Instruction *I = dyn_cast<Instruction>(V))
270 if (isInstructionTriviallyDead(I, TLI))
271 deleteDeadInstruction(I, SE, TLI);
274 //===----------------------------------------------------------------------===//
276 // Implementation of LIRUtil
278 //===----------------------------------------------------------------------===//
280 // This function will return true iff the given block contains nothing but goto.
281 // A typical usage of this function is to check if the preheader function is
282 // "almost" empty such that generated intrinsic functions can be moved across
283 // the preheader and be placed at the end of the precondition block without
284 // the concern of breaking data dependence.
285 bool LIRUtil::isAlmostEmpty(BasicBlock *BB) {
286 if (BranchInst *Br = getBranch(BB)) {
287 return Br->isUnconditional() && BB->size() == 1;
292 BasicBlock *LIRUtil::getPrecondBb(BasicBlock *PreHead) {
293 if (BasicBlock *BB = PreHead->getSinglePredecessor()) {
294 BranchInst *Br = getBranch(BB);
295 return Br && Br->isConditional() ? BB : 0;
300 //===----------------------------------------------------------------------===//
302 // Implementation of NclPopcountRecognize
304 //===----------------------------------------------------------------------===//
306 NclPopcountRecognize::NclPopcountRecognize(LoopIdiomRecognize &TheLIR):
307 LIR(TheLIR), CurLoop(TheLIR.getLoop()), PreCondBB(0) {
310 bool NclPopcountRecognize::preliminaryScreen() {
311 const TargetTransformInfo *TTI = LIR.getTargetTransformInfo();
312 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
315 // Counting population are usually conducted by few arithmetic instructions.
316 // Such instructions can be easilly "absorbed" by vacant slots in a
317 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
318 // in a compact loop.
320 // Give up if the loop has multiple blocks or multiple backedges.
321 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
324 BasicBlock *LoopBody = *(CurLoop->block_begin());
325 if (LoopBody->size() >= 20) {
326 // The loop is too big, bail out.
330 // It should have a preheader containing nothing but a goto instruction.
331 BasicBlock *PreHead = CurLoop->getLoopPreheader();
332 if (!PreHead || !LIRUtil::isAlmostEmpty(PreHead))
335 // It should have a precondition block where the generated popcount instrinsic
336 // function will be inserted.
337 PreCondBB = LIRUtil::getPrecondBb(PreHead);
344 Value *NclPopcountRecognize::matchCondition (BranchInst *Br,
345 BasicBlock *LoopEntry) const {
346 if (!Br || !Br->isConditional())
349 ICmpInst *Cond = dyn_cast<ICmpInst>(Br->getCondition());
353 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
354 if (!CmpZero || !CmpZero->isZero())
357 ICmpInst::Predicate Pred = Cond->getPredicate();
358 if ((Pred == ICmpInst::ICMP_NE && Br->getSuccessor(0) == LoopEntry) ||
359 (Pred == ICmpInst::ICMP_EQ && Br->getSuccessor(1) == LoopEntry))
360 return Cond->getOperand(0);
365 bool NclPopcountRecognize::detectIdiom(Instruction *&CntInst,
368 // Following code tries to detect this idiom:
371 // goto loop-exit // the precondition of the loop
374 // x1 = phi (x0, x2);
375 // cnt1 = phi(cnt0, cnt2);
379 // x2 = x1 & (x1 - 1);
386 // step 1: Check to see if the look-back branch match this pattern:
387 // "if (a!=0) goto loop-entry".
388 BasicBlock *LoopEntry;
389 Instruction *DefX2, *CountInst;
390 Value *VarX1, *VarX0;
391 PHINode *PhiX, *CountPhi;
393 DefX2 = CountInst = 0;
396 LoopEntry = *(CurLoop->block_begin());
398 // step 1: Check if the loop-back branch is in desirable form.
400 if (Value *T = matchCondition (LIRUtil::getBranch(LoopEntry), LoopEntry))
401 DefX2 = dyn_cast<Instruction>(T);
406 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
408 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
411 BinaryOperator *SubOneOp;
413 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
414 VarX1 = DefX2->getOperand(1);
416 VarX1 = DefX2->getOperand(0);
417 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
422 Instruction *SubInst = cast<Instruction>(SubOneOp);
423 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
425 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
426 (SubInst->getOpcode() == Instruction::Add && Dec->isAllOnesValue()))) {
431 // step 3: Check the recurrence of variable X
433 PhiX = dyn_cast<PHINode>(VarX1);
435 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
440 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
443 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI(),
444 IterE = LoopEntry->end(); Iter != IterE; Iter++) {
445 Instruction *Inst = Iter;
446 if (Inst->getOpcode() != Instruction::Add)
449 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
450 if (!Inc || !Inc->isOne())
453 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
454 if (!Phi || Phi->getParent() != LoopEntry)
457 // Check if the result of the instruction is live of the loop.
458 bool LiveOutLoop = false;
459 for (User *U : Inst->users()) {
460 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
461 LiveOutLoop = true; break;
476 // step 5: check if the precondition is in this form:
477 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
479 BranchInst *PreCondBr = LIRUtil::getBranch(PreCondBB);
480 Value *T = matchCondition (PreCondBr, CurLoop->getLoopPreheader());
481 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
492 void NclPopcountRecognize::transform(Instruction *CntInst,
493 PHINode *CntPhi, Value *Var) {
495 ScalarEvolution *SE = LIR.getScalarEvolution();
496 TargetLibraryInfo *TLI = LIR.getTargetLibraryInfo();
497 BasicBlock *PreHead = CurLoop->getLoopPreheader();
498 BranchInst *PreCondBr = LIRUtil::getBranch(PreCondBB);
499 const DebugLoc DL = CntInst->getDebugLoc();
501 // Assuming before transformation, the loop is following:
502 // if (x) // the precondition
503 // do { cnt++; x &= x - 1; } while(x);
505 // Step 1: Insert the ctpop instruction at the end of the precondition block
506 IRBuilderTy Builder(PreCondBr);
507 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
509 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
510 NewCount = PopCntZext =
511 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
513 if (NewCount != PopCnt)
514 (cast<Instruction>(NewCount))->setDebugLoc(DL);
516 // TripCnt is exactly the number of iterations the loop has
519 // If the population counter's initial value is not zero, insert Add Inst.
520 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
521 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
522 if (!InitConst || !InitConst->isZero()) {
523 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
524 (cast<Instruction>(NewCount))->setDebugLoc(DL);
528 // Step 2: Replace the precondition from "if(x == 0) goto loop-exit" to
529 // "if(NewCount == 0) loop-exit". Withtout this change, the intrinsic
530 // function would be partial dead code, and downstream passes will drag
531 // it back from the precondition block to the preheader.
533 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
535 Value *Opnd0 = PopCntZext;
536 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
537 if (PreCond->getOperand(0) != Var)
538 std::swap(Opnd0, Opnd1);
540 ICmpInst *NewPreCond =
541 cast<ICmpInst>(Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
542 PreCond->replaceAllUsesWith(NewPreCond);
544 deleteDeadInstruction(PreCond, *SE, TLI);
547 // Step 3: Note that the population count is exactly the trip count of the
548 // loop in question, which enble us to to convert the loop from noncountable
549 // loop into a countable one. The benefit is twofold:
551 // - If the loop only counts population, the entire loop become dead after
552 // the transformation. It is lots easier to prove a countable loop dead
553 // than to prove a noncountable one. (In some C dialects, a infite loop
554 // isn't dead even if it computes nothing useful. In general, DCE needs
555 // to prove a noncountable loop finite before safely delete it.)
557 // - If the loop also performs something else, it remains alive.
558 // Since it is transformed to countable form, it can be aggressively
559 // optimized by some optimizations which are in general not applicable
560 // to a noncountable loop.
562 // After this step, this loop (conceptually) would look like following:
563 // newcnt = __builtin_ctpop(x);
566 // do { cnt++; x &= x-1; t--) } while (t > 0);
567 BasicBlock *Body = *(CurLoop->block_begin());
569 BranchInst *LbBr = LIRUtil::getBranch(Body);
570 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
571 Type *Ty = TripCnt->getType();
573 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", Body->begin());
575 Builder.SetInsertPoint(LbCond);
576 Value *Opnd1 = cast<Value>(TcPhi);
577 Value *Opnd2 = cast<Value>(ConstantInt::get(Ty, 1));
579 cast<Instruction>(Builder.CreateSub(Opnd1, Opnd2, "tcdec", false, true));
581 TcPhi->addIncoming(TripCnt, PreHead);
582 TcPhi->addIncoming(TcDec, Body);
584 CmpInst::Predicate Pred = (LbBr->getSuccessor(0) == Body) ?
585 CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
586 LbCond->setPredicate(Pred);
587 LbCond->setOperand(0, TcDec);
588 LbCond->setOperand(1, cast<Value>(ConstantInt::get(Ty, 0)));
591 // Step 4: All the references to the original population counter outside
592 // the loop are replaced with the NewCount -- the value returned from
593 // __builtin_ctpop().
595 SmallVector<Value *, 4> CntUses;
596 for (User *U : CntInst->users())
597 if (cast<Instruction>(U)->getParent() != Body)
598 CntUses.push_back(U);
599 for (unsigned Idx = 0; Idx < CntUses.size(); Idx++) {
600 (cast<Instruction>(CntUses[Idx]))->replaceUsesOfWith(CntInst, NewCount);
604 // step 5: Forget the "non-computable" trip-count SCEV associated with the
605 // loop. The loop would otherwise not be deleted even if it becomes empty.
606 SE->forgetLoop(CurLoop);
609 CallInst *NclPopcountRecognize::createPopcntIntrinsic(IRBuilderTy &IRBuilder,
610 Value *Val, DebugLoc DL) {
611 Value *Ops[] = { Val };
612 Type *Tys[] = { Val->getType() };
614 Module *M = (*(CurLoop->block_begin()))->getParent()->getParent();
615 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
616 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
622 /// recognize - detect population count idiom in a non-countable loop. If
623 /// detected, transform the relevant code to popcount intrinsic function
624 /// call, and return true; otherwise, return false.
625 bool NclPopcountRecognize::recognize() {
627 if (!LIR.getTargetTransformInfo())
630 LIR.getScalarEvolution();
632 if (!preliminaryScreen())
635 Instruction *CntInst;
638 if (!detectIdiom(CntInst, CntPhi, Val))
641 transform(CntInst, CntPhi, Val);
645 //===----------------------------------------------------------------------===//
647 // Implementation of LoopIdiomRecognize
649 //===----------------------------------------------------------------------===//
651 bool LoopIdiomRecognize::runOnCountableLoop() {
652 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
653 if (isa<SCEVCouldNotCompute>(BECount)) return false;
655 // If this loop executes exactly one time, then it should be peeled, not
656 // optimized by this pass.
657 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
658 if (BECst->getValue()->getValue() == 0)
661 // We require target data for now.
662 if (!getDataLayout())
666 (void)getDominatorTree();
668 LoopInfo &LI = getAnalysis<LoopInfo>();
669 TLI = &getAnalysis<TargetLibraryInfo>();
672 (void)getTargetLibraryInfo();
674 SmallVector<BasicBlock*, 8> ExitBlocks;
675 CurLoop->getUniqueExitBlocks(ExitBlocks);
677 DEBUG(dbgs() << "loop-idiom Scanning: F["
678 << CurLoop->getHeader()->getParent()->getName()
679 << "] Loop %" << CurLoop->getHeader()->getName() << "\n");
681 bool MadeChange = false;
682 // Scan all the blocks in the loop that are not in subloops.
683 for (Loop::block_iterator BI = CurLoop->block_begin(),
684 E = CurLoop->block_end(); BI != E; ++BI) {
685 // Ignore blocks in subloops.
686 if (LI.getLoopFor(*BI) != CurLoop)
689 MadeChange |= runOnLoopBlock(*BI, BECount, ExitBlocks);
694 bool LoopIdiomRecognize::runOnNoncountableLoop() {
695 NclPopcountRecognize Popcount(*this);
696 if (Popcount.recognize())
702 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
703 if (skipOptnoneFunction(L))
708 // If the loop could not be converted to canonical form, it must have an
709 // indirectbr in it, just give up.
710 if (!L->getLoopPreheader())
713 // Disable loop idiom recognition if the function's name is a common idiom.
714 StringRef Name = L->getHeader()->getParent()->getName();
715 if (Name == "memset" || Name == "memcpy")
718 SE = &getAnalysis<ScalarEvolution>();
719 if (SE->hasLoopInvariantBackedgeTakenCount(L))
720 return runOnCountableLoop();
721 return runOnNoncountableLoop();
724 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
725 /// with the specified backedge count. This block is known to be in the current
726 /// loop and not in any subloops.
727 bool LoopIdiomRecognize::runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
728 SmallVectorImpl<BasicBlock*> &ExitBlocks) {
729 // We can only promote stores in this block if they are unconditionally
730 // executed in the loop. For a block to be unconditionally executed, it has
731 // to dominate all the exit blocks of the loop. Verify this now.
732 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
733 if (!DT->dominates(BB, ExitBlocks[i]))
736 bool MadeChange = false;
737 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
738 Instruction *Inst = I++;
739 // Look for store instructions, which may be optimized to memset/memcpy.
740 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
742 if (!processLoopStore(SI, BECount)) continue;
745 // If processing the store invalidated our iterator, start over from the
752 // Look for memset instructions, which may be optimized to a larger memset.
753 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
755 if (!processLoopMemSet(MSI, BECount)) continue;
758 // If processing the memset invalidated our iterator, start over from the
770 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
771 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
772 if (!SI->isSimple()) return false;
774 Value *StoredVal = SI->getValueOperand();
775 Value *StorePtr = SI->getPointerOperand();
777 // Reject stores that are so large that they overflow an unsigned.
778 uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
779 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
782 // See if the pointer expression is an AddRec like {base,+,1} on the current
783 // loop, which indicates a strided store. If we have something else, it's a
784 // random store we can't handle.
785 const SCEVAddRecExpr *StoreEv =
786 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
787 if (StoreEv == 0 || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
790 // Check to see if the stride matches the size of the store. If so, then we
791 // know that every byte is touched in the loop.
792 unsigned StoreSize = (unsigned)SizeInBits >> 3;
793 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
795 if (Stride == 0 || StoreSize != Stride->getValue()->getValue()) {
796 // TODO: Could also handle negative stride here someday, that will require
797 // the validity check in mayLoopAccessLocation to be updated though.
798 // Enable this to print exact negative strides.
799 if (0 && Stride && StoreSize == -Stride->getValue()->getValue()) {
800 dbgs() << "NEGATIVE STRIDE: " << *SI << "\n";
801 dbgs() << "BB: " << *SI->getParent();
807 // See if we can optimize just this store in isolation.
808 if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
809 StoredVal, SI, StoreEv, BECount))
812 // If the stored value is a strided load in the same loop with the same stride
813 // this this may be transformable into a memcpy. This kicks in for stuff like
814 // for (i) A[i] = B[i];
815 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
816 const SCEVAddRecExpr *LoadEv =
817 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
818 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
819 StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple())
820 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
823 //errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
828 /// processLoopMemSet - See if this memset can be promoted to a large memset.
829 bool LoopIdiomRecognize::
830 processLoopMemSet(MemSetInst *MSI, const SCEV *BECount) {
831 // We can only handle non-volatile memsets with a constant size.
832 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength())) return false;
834 // If we're not allowed to hack on memset, we fail.
835 if (!TLI->has(LibFunc::memset))
838 Value *Pointer = MSI->getDest();
840 // See if the pointer expression is an AddRec like {base,+,1} on the current
841 // loop, which indicates a strided store. If we have something else, it's a
842 // random store we can't handle.
843 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
844 if (Ev == 0 || Ev->getLoop() != CurLoop || !Ev->isAffine())
847 // Reject memsets that are so large that they overflow an unsigned.
848 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
849 if ((SizeInBytes >> 32) != 0)
852 // Check to see if the stride matches the size of the memset. If so, then we
853 // know that every byte is touched in the loop.
854 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
856 // TODO: Could also handle negative stride here someday, that will require the
857 // validity check in mayLoopAccessLocation to be updated though.
858 if (Stride == 0 || MSI->getLength() != Stride->getValue())
861 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
862 MSI->getAlignment(), MSI->getValue(),
867 /// mayLoopAccessLocation - Return true if the specified loop might access the
868 /// specified pointer location, which is a loop-strided access. The 'Access'
869 /// argument specifies what the verboten forms of access are (read or write).
870 static bool mayLoopAccessLocation(Value *Ptr,AliasAnalysis::ModRefResult Access,
871 Loop *L, const SCEV *BECount,
872 unsigned StoreSize, AliasAnalysis &AA,
873 Instruction *IgnoredStore) {
874 // Get the location that may be stored across the loop. Since the access is
875 // strided positively through memory, we say that the modified location starts
876 // at the pointer and has infinite size.
877 uint64_t AccessSize = AliasAnalysis::UnknownSize;
879 // If the loop iterates a fixed number of times, we can refine the access size
880 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
881 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
882 AccessSize = (BECst->getValue()->getZExtValue()+1)*StoreSize;
884 // TODO: For this to be really effective, we have to dive into the pointer
885 // operand in the store. Store to &A[i] of 100 will always return may alias
886 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
887 // which will then no-alias a store to &A[100].
888 AliasAnalysis::Location StoreLoc(Ptr, AccessSize);
890 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
892 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
893 if (&*I != IgnoredStore &&
894 (AA.getModRefInfo(I, StoreLoc) & Access))
900 /// getMemSetPatternValue - If a strided store of the specified value is safe to
901 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
902 /// be passed in. Otherwise, return null.
904 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
905 /// just replicate their input array and then pass on to memset_pattern16.
906 static Constant *getMemSetPatternValue(Value *V, const DataLayout &DL) {
907 // If the value isn't a constant, we can't promote it to being in a constant
908 // array. We could theoretically do a store to an alloca or something, but
909 // that doesn't seem worthwhile.
910 Constant *C = dyn_cast<Constant>(V);
911 if (C == 0) return 0;
913 // Only handle simple values that are a power of two bytes in size.
914 uint64_t Size = DL.getTypeSizeInBits(V->getType());
915 if (Size == 0 || (Size & 7) || (Size & (Size-1)))
918 // Don't care enough about darwin/ppc to implement this.
919 if (DL.isBigEndian())
922 // Convert to size in bytes.
925 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
926 // if the top and bottom are the same (e.g. for vectors and large integers).
927 if (Size > 16) return 0;
929 // If the constant is exactly 16 bytes, just use it.
930 if (Size == 16) return C;
932 // Otherwise, we'll use an array of the constants.
933 unsigned ArraySize = 16/Size;
934 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
935 return ConstantArray::get(AT, std::vector<Constant*>(ArraySize, C));
939 /// processLoopStridedStore - We see a strided store of some value. If we can
940 /// transform this into a memset or memset_pattern in the loop preheader, do so.
941 bool LoopIdiomRecognize::
942 processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
943 unsigned StoreAlignment, Value *StoredVal,
944 Instruction *TheStore, const SCEVAddRecExpr *Ev,
945 const SCEV *BECount) {
947 // If the stored value is a byte-wise value (like i32 -1), then it may be
948 // turned into a memset of i8 -1, assuming that all the consecutive bytes
949 // are stored. A store of i32 0x01020304 can never be turned into a memset,
950 // but it can be turned into memset_pattern if the target supports it.
951 Value *SplatValue = isBytewiseValue(StoredVal);
952 Constant *PatternValue = 0;
954 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
956 // If we're allowed to form a memset, and the stored value would be acceptable
957 // for memset, use it.
958 if (SplatValue && TLI->has(LibFunc::memset) &&
959 // Verify that the stored value is loop invariant. If not, we can't
960 // promote the memset.
961 CurLoop->isLoopInvariant(SplatValue)) {
962 // Keep and use SplatValue.
964 } else if (DestAS == 0 &&
965 TLI->has(LibFunc::memset_pattern16) &&
966 (PatternValue = getMemSetPatternValue(StoredVal, *DL))) {
967 // Don't create memset_pattern16s with address spaces.
968 // It looks like we can use PatternValue!
971 // Otherwise, this isn't an idiom we can transform. For example, we can't
972 // do anything with a 3-byte store.
976 // The trip count of the loop and the base pointer of the addrec SCEV is
977 // guaranteed to be loop invariant, which means that it should dominate the
978 // header. This allows us to insert code for it in the preheader.
979 BasicBlock *Preheader = CurLoop->getLoopPreheader();
980 IRBuilder<> Builder(Preheader->getTerminator());
981 SCEVExpander Expander(*SE, "loop-idiom");
983 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
985 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
986 // this into a memset in the loop preheader now if we want. However, this
987 // would be unsafe to do if there is anything else in the loop that may read
988 // or write to the aliased location. Check for any overlap by generating the
989 // base pointer and checking the region.
991 Expander.expandCodeFor(Ev->getStart(), DestInt8PtrTy,
992 Preheader->getTerminator());
994 if (mayLoopAccessLocation(BasePtr, AliasAnalysis::ModRef,
996 StoreSize, getAnalysis<AliasAnalysis>(), TheStore)) {
998 // If we generated new code for the base pointer, clean up.
999 deleteIfDeadInstruction(BasePtr, *SE, TLI);
1003 // Okay, everything looks good, insert the memset.
1005 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
1006 // pointer size if it isn't already.
1007 Type *IntPtr = Builder.getIntPtrTy(DL, DestAS);
1008 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
1010 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
1012 if (StoreSize != 1) {
1013 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
1018 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
1022 NewCall = Builder.CreateMemSet(BasePtr,
1027 // Everything is emitted in default address space
1028 Type *Int8PtrTy = DestInt8PtrTy;
1030 Module *M = TheStore->getParent()->getParent()->getParent();
1031 Value *MSP = M->getOrInsertFunction("memset_pattern16",
1032 Builder.getVoidTy(),
1038 // Otherwise we should form a memset_pattern16. PatternValue is known to be
1039 // an constant array of 16-bytes. Plop the value into a mergable global.
1040 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
1041 GlobalValue::InternalLinkage,
1042 PatternValue, ".memset_pattern");
1043 GV->setUnnamedAddr(true); // Ok to merge these.
1044 GV->setAlignment(16);
1045 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
1046 NewCall = Builder.CreateCall3(MSP, BasePtr, PatternPtr, NumBytes);
1049 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
1050 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
1051 NewCall->setDebugLoc(TheStore->getDebugLoc());
1053 // Okay, the memset has been formed. Zap the original store and anything that
1055 deleteDeadInstruction(TheStore, *SE, TLI);
1060 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
1061 /// same-strided load.
1062 bool LoopIdiomRecognize::
1063 processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
1064 const SCEVAddRecExpr *StoreEv,
1065 const SCEVAddRecExpr *LoadEv,
1066 const SCEV *BECount) {
1067 // If we're not allowed to form memcpy, we fail.
1068 if (!TLI->has(LibFunc::memcpy))
1071 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
1073 // The trip count of the loop and the base pointer of the addrec SCEV is
1074 // guaranteed to be loop invariant, which means that it should dominate the
1075 // header. This allows us to insert code for it in the preheader.
1076 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1077 IRBuilder<> Builder(Preheader->getTerminator());
1078 SCEVExpander Expander(*SE, "loop-idiom");
1080 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
1081 // this into a memcpy in the loop preheader now if we want. However, this
1082 // would be unsafe to do if there is anything else in the loop that may read
1083 // or write the memory region we're storing to. This includes the load that
1084 // feeds the stores. Check for an alias by generating the base address and
1085 // checking everything.
1086 Value *StoreBasePtr =
1087 Expander.expandCodeFor(StoreEv->getStart(),
1088 Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
1089 Preheader->getTerminator());
1091 if (mayLoopAccessLocation(StoreBasePtr, AliasAnalysis::ModRef,
1092 CurLoop, BECount, StoreSize,
1093 getAnalysis<AliasAnalysis>(), SI)) {
1095 // If we generated new code for the base pointer, clean up.
1096 deleteIfDeadInstruction(StoreBasePtr, *SE, TLI);
1100 // For a memcpy, we have to make sure that the input array is not being
1101 // mutated by the loop.
1102 Value *LoadBasePtr =
1103 Expander.expandCodeFor(LoadEv->getStart(),
1104 Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
1105 Preheader->getTerminator());
1107 if (mayLoopAccessLocation(LoadBasePtr, AliasAnalysis::Mod, CurLoop, BECount,
1108 StoreSize, getAnalysis<AliasAnalysis>(), SI)) {
1110 // If we generated new code for the base pointer, clean up.
1111 deleteIfDeadInstruction(LoadBasePtr, *SE, TLI);
1112 deleteIfDeadInstruction(StoreBasePtr, *SE, TLI);
1116 // Okay, everything is safe, we can transform this!
1119 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
1120 // pointer size if it isn't already.
1121 Type *IntPtrTy = Builder.getIntPtrTy(DL, SI->getPointerAddressSpace());
1122 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
1124 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtrTy, 1),
1127 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
1131 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
1134 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
1135 std::min(SI->getAlignment(), LI->getAlignment()));
1136 NewCall->setDebugLoc(SI->getDebugLoc());
1138 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
1139 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
1140 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
1143 // Okay, the memset has been formed. Zap the original store and anything that
1145 deleteDeadInstruction(SI, *SE, TLI);