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 #include "llvm/Transforms/Scalar.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/Analysis/AliasAnalysis.h"
47 #include "llvm/Analysis/LoopPass.h"
48 #include "llvm/Analysis/ScalarEvolutionExpander.h"
49 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
50 #include "llvm/Analysis/TargetTransformInfo.h"
51 #include "llvm/Analysis/ValueTracking.h"
52 #include "llvm/IR/DataLayout.h"
53 #include "llvm/IR/Dominators.h"
54 #include "llvm/IR/IRBuilder.h"
55 #include "llvm/IR/IntrinsicInst.h"
56 #include "llvm/IR/Module.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include "llvm/Analysis/TargetLibraryInfo.h"
60 #include "llvm/Transforms/Utils/Local.h"
63 #define DEBUG_TYPE "loop-idiom"
65 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
66 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
70 class LoopIdiomRecognize;
72 /// This class defines some utility functions for loop idiom recognization.
75 /// Return true iff the block contains nothing but an uncondition branch
76 /// (aka goto instruction).
77 static bool isAlmostEmpty(BasicBlock *);
79 static BranchInst *getBranch(BasicBlock *BB) {
80 return dyn_cast<BranchInst>(BB->getTerminator());
83 /// Derive the precondition block (i.e the block that guards the loop
84 /// preheader) from the given preheader.
85 static BasicBlock *getPrecondBb(BasicBlock *PreHead);
88 /// This class is to recoginize idioms of population-count conducted in
89 /// a noncountable loop. Currently it only recognizes this pattern:
91 /// while(x) {cnt++; ...; x &= x - 1; ...}
93 class NclPopcountRecognize {
94 LoopIdiomRecognize &LIR;
96 BasicBlock *PreCondBB;
98 typedef IRBuilder<> IRBuilderTy;
101 explicit NclPopcountRecognize(LoopIdiomRecognize &TheLIR);
105 /// Take a glimpse of the loop to see if we need to go ahead recoginizing
107 bool preliminaryScreen();
109 /// Check if the given conditional branch is based on the comparison
110 /// between a variable and zero, and if the variable is non-zero, the
111 /// control yields to the loop entry. If the branch matches the behavior,
112 /// the variable involved in the comparion is returned. This function will
113 /// be called to see if the precondition and postcondition of the loop
114 /// are in desirable form.
115 Value *matchCondition(BranchInst *Br, BasicBlock *NonZeroTarget) const;
117 /// Return true iff the idiom is detected in the loop. and 1) \p CntInst
118 /// is set to the instruction counting the population bit. 2) \p CntPhi
119 /// is set to the corresponding phi node. 3) \p Var is set to the value
120 /// whose population bits are being counted.
122 (Instruction *&CntInst, PHINode *&CntPhi, Value *&Var) const;
124 /// Insert ctpop intrinsic function and some obviously dead instructions.
125 void transform(Instruction *CntInst, PHINode *CntPhi, Value *Var);
127 /// Create llvm.ctpop.* intrinsic function.
128 CallInst *createPopcntIntrinsic(IRBuilderTy &IRB, Value *Val, DebugLoc DL);
131 class LoopIdiomRecognize : public LoopPass {
133 const DataLayout *DL;
136 TargetLibraryInfo *TLI;
137 const TargetTransformInfo *TTI;
140 explicit LoopIdiomRecognize() : LoopPass(ID) {
141 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
142 DL = nullptr; DT = nullptr; SE = nullptr; TLI = nullptr; TTI = nullptr;
145 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
146 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
147 SmallVectorImpl<BasicBlock*> &ExitBlocks);
149 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
150 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
152 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
153 unsigned StoreAlignment,
154 Value *SplatValue, Instruction *TheStore,
155 const SCEVAddRecExpr *Ev,
156 const SCEV *BECount);
157 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
158 const SCEVAddRecExpr *StoreEv,
159 const SCEVAddRecExpr *LoadEv,
160 const SCEV *BECount);
162 /// This transformation requires natural loop information & requires that
163 /// loop preheaders be inserted into the CFG.
165 void getAnalysisUsage(AnalysisUsage &AU) const override {
166 AU.addRequired<LoopInfoWrapperPass>();
167 AU.addPreserved<LoopInfoWrapperPass>();
168 AU.addRequiredID(LoopSimplifyID);
169 AU.addPreservedID(LoopSimplifyID);
170 AU.addRequiredID(LCSSAID);
171 AU.addPreservedID(LCSSAID);
172 AU.addRequired<AliasAnalysis>();
173 AU.addPreserved<AliasAnalysis>();
174 AU.addRequired<ScalarEvolution>();
175 AU.addPreserved<ScalarEvolution>();
176 AU.addPreserved<DominatorTreeWrapperPass>();
177 AU.addRequired<DominatorTreeWrapperPass>();
178 AU.addRequired<TargetLibraryInfoWrapperPass>();
179 AU.addRequired<TargetTransformInfoWrapperPass>();
182 const DataLayout *getDataLayout() {
185 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
186 DL = DLP ? &DLP->getDataLayout() : nullptr;
190 DominatorTree *getDominatorTree() {
192 : (DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree());
195 ScalarEvolution *getScalarEvolution() {
196 return SE ? SE : (SE = &getAnalysis<ScalarEvolution>());
199 TargetLibraryInfo *getTargetLibraryInfo() {
201 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
206 const TargetTransformInfo *getTargetTransformInfo() {
208 : (TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
209 *CurLoop->getHeader()->getParent()));
212 Loop *getLoop() const { return CurLoop; }
215 bool runOnNoncountableLoop();
216 bool runOnCountableLoop();
220 char LoopIdiomRecognize::ID = 0;
221 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
223 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
224 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
225 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
226 INITIALIZE_PASS_DEPENDENCY(LCSSA)
227 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
228 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
229 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
230 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
231 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
234 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
236 /// deleteDeadInstruction - Delete this instruction. Before we do, go through
237 /// and zero out all the operands of this instruction. If any of them become
238 /// dead, delete them and the computation tree that feeds them.
240 static void deleteDeadInstruction(Instruction *I,
241 const TargetLibraryInfo *TLI) {
242 SmallVector<Value *, 16> Operands(I->value_op_begin(), I->value_op_end());
243 I->replaceAllUsesWith(UndefValue::get(I->getType()));
244 I->eraseFromParent();
245 for (Value *Op : Operands)
246 RecursivelyDeleteTriviallyDeadInstructions(Op, TLI);
249 //===----------------------------------------------------------------------===//
251 // Implementation of LIRUtil
253 //===----------------------------------------------------------------------===//
255 // This function will return true iff the given block contains nothing but goto.
256 // A typical usage of this function is to check if the preheader function is
257 // "almost" empty such that generated intrinsic functions can be moved across
258 // the preheader and be placed at the end of the precondition block without
259 // the concern of breaking data dependence.
260 bool LIRUtil::isAlmostEmpty(BasicBlock *BB) {
261 if (BranchInst *Br = getBranch(BB)) {
262 return Br->isUnconditional() && Br == BB->begin();
267 BasicBlock *LIRUtil::getPrecondBb(BasicBlock *PreHead) {
268 if (BasicBlock *BB = PreHead->getSinglePredecessor()) {
269 BranchInst *Br = getBranch(BB);
270 return Br && Br->isConditional() ? BB : nullptr;
275 //===----------------------------------------------------------------------===//
277 // Implementation of NclPopcountRecognize
279 //===----------------------------------------------------------------------===//
281 NclPopcountRecognize::NclPopcountRecognize(LoopIdiomRecognize &TheLIR):
282 LIR(TheLIR), CurLoop(TheLIR.getLoop()), PreCondBB(nullptr) {
285 bool NclPopcountRecognize::preliminaryScreen() {
286 const TargetTransformInfo *TTI = LIR.getTargetTransformInfo();
287 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
290 // Counting population are usually conducted by few arithmetic instructions.
291 // Such instructions can be easilly "absorbed" by vacant slots in a
292 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
293 // in a compact loop.
295 // Give up if the loop has multiple blocks or multiple backedges.
296 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
299 BasicBlock *LoopBody = *(CurLoop->block_begin());
300 if (LoopBody->size() >= 20) {
301 // The loop is too big, bail out.
305 // It should have a preheader containing nothing but a goto instruction.
306 BasicBlock *PreHead = CurLoop->getLoopPreheader();
307 if (!PreHead || !LIRUtil::isAlmostEmpty(PreHead))
310 // It should have a precondition block where the generated popcount instrinsic
311 // function will be inserted.
312 PreCondBB = LIRUtil::getPrecondBb(PreHead);
319 Value *NclPopcountRecognize::matchCondition(BranchInst *Br,
320 BasicBlock *LoopEntry) const {
321 if (!Br || !Br->isConditional())
324 ICmpInst *Cond = dyn_cast<ICmpInst>(Br->getCondition());
328 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
329 if (!CmpZero || !CmpZero->isZero())
332 ICmpInst::Predicate Pred = Cond->getPredicate();
333 if ((Pred == ICmpInst::ICMP_NE && Br->getSuccessor(0) == LoopEntry) ||
334 (Pred == ICmpInst::ICMP_EQ && Br->getSuccessor(1) == LoopEntry))
335 return Cond->getOperand(0);
340 bool NclPopcountRecognize::detectIdiom(Instruction *&CntInst,
343 // Following code tries to detect this idiom:
346 // goto loop-exit // the precondition of the loop
349 // x1 = phi (x0, x2);
350 // cnt1 = phi(cnt0, cnt2);
354 // x2 = x1 & (x1 - 1);
361 // step 1: Check to see if the look-back branch match this pattern:
362 // "if (a!=0) goto loop-entry".
363 BasicBlock *LoopEntry;
364 Instruction *DefX2, *CountInst;
365 Value *VarX1, *VarX0;
366 PHINode *PhiX, *CountPhi;
368 DefX2 = CountInst = nullptr;
369 VarX1 = VarX0 = nullptr;
370 PhiX = CountPhi = nullptr;
371 LoopEntry = *(CurLoop->block_begin());
373 // step 1: Check if the loop-back branch is in desirable form.
375 if (Value *T = matchCondition (LIRUtil::getBranch(LoopEntry), LoopEntry))
376 DefX2 = dyn_cast<Instruction>(T);
381 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
383 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
386 BinaryOperator *SubOneOp;
388 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
389 VarX1 = DefX2->getOperand(1);
391 VarX1 = DefX2->getOperand(0);
392 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
397 Instruction *SubInst = cast<Instruction>(SubOneOp);
398 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
400 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
401 (SubInst->getOpcode() == Instruction::Add && Dec->isAllOnesValue()))) {
406 // step 3: Check the recurrence of variable X
408 PhiX = dyn_cast<PHINode>(VarX1);
410 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
415 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
418 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI(),
419 IterE = LoopEntry->end(); Iter != IterE; Iter++) {
420 Instruction *Inst = Iter;
421 if (Inst->getOpcode() != Instruction::Add)
424 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
425 if (!Inc || !Inc->isOne())
428 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
429 if (!Phi || Phi->getParent() != LoopEntry)
432 // Check if the result of the instruction is live of the loop.
433 bool LiveOutLoop = false;
434 for (User *U : Inst->users()) {
435 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
436 LiveOutLoop = true; break;
451 // step 5: check if the precondition is in this form:
452 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
454 BranchInst *PreCondBr = LIRUtil::getBranch(PreCondBB);
455 Value *T = matchCondition (PreCondBr, CurLoop->getLoopPreheader());
456 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
467 void NclPopcountRecognize::transform(Instruction *CntInst,
468 PHINode *CntPhi, Value *Var) {
470 ScalarEvolution *SE = LIR.getScalarEvolution();
471 TargetLibraryInfo *TLI = LIR.getTargetLibraryInfo();
472 BasicBlock *PreHead = CurLoop->getLoopPreheader();
473 BranchInst *PreCondBr = LIRUtil::getBranch(PreCondBB);
474 const DebugLoc DL = CntInst->getDebugLoc();
476 // Assuming before transformation, the loop is following:
477 // if (x) // the precondition
478 // do { cnt++; x &= x - 1; } while(x);
480 // Step 1: Insert the ctpop instruction at the end of the precondition block
481 IRBuilderTy Builder(PreCondBr);
482 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
484 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
485 NewCount = PopCntZext =
486 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
488 if (NewCount != PopCnt)
489 (cast<Instruction>(NewCount))->setDebugLoc(DL);
491 // TripCnt is exactly the number of iterations the loop has
494 // If the population counter's initial value is not zero, insert Add Inst.
495 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
496 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
497 if (!InitConst || !InitConst->isZero()) {
498 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
499 (cast<Instruction>(NewCount))->setDebugLoc(DL);
503 // Step 2: Replace the precondition from "if(x == 0) goto loop-exit" to
504 // "if(NewCount == 0) loop-exit". Withtout this change, the intrinsic
505 // function would be partial dead code, and downstream passes will drag
506 // it back from the precondition block to the preheader.
508 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
510 Value *Opnd0 = PopCntZext;
511 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
512 if (PreCond->getOperand(0) != Var)
513 std::swap(Opnd0, Opnd1);
515 ICmpInst *NewPreCond =
516 cast<ICmpInst>(Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
517 PreCond->replaceAllUsesWith(NewPreCond);
519 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
522 // Step 3: Note that the population count is exactly the trip count of the
523 // loop in question, which enble us to to convert the loop from noncountable
524 // loop into a countable one. The benefit is twofold:
526 // - If the loop only counts population, the entire loop become dead after
527 // the transformation. It is lots easier to prove a countable loop dead
528 // than to prove a noncountable one. (In some C dialects, a infite loop
529 // isn't dead even if it computes nothing useful. In general, DCE needs
530 // to prove a noncountable loop finite before safely delete it.)
532 // - If the loop also performs something else, it remains alive.
533 // Since it is transformed to countable form, it can be aggressively
534 // optimized by some optimizations which are in general not applicable
535 // to a noncountable loop.
537 // After this step, this loop (conceptually) would look like following:
538 // newcnt = __builtin_ctpop(x);
541 // do { cnt++; x &= x-1; t--) } while (t > 0);
542 BasicBlock *Body = *(CurLoop->block_begin());
544 BranchInst *LbBr = LIRUtil::getBranch(Body);
545 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
546 Type *Ty = TripCnt->getType();
548 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", Body->begin());
550 Builder.SetInsertPoint(LbCond);
551 Value *Opnd1 = cast<Value>(TcPhi);
552 Value *Opnd2 = cast<Value>(ConstantInt::get(Ty, 1));
554 cast<Instruction>(Builder.CreateSub(Opnd1, Opnd2, "tcdec", false, true));
556 TcPhi->addIncoming(TripCnt, PreHead);
557 TcPhi->addIncoming(TcDec, Body);
559 CmpInst::Predicate Pred = (LbBr->getSuccessor(0) == Body) ?
560 CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
561 LbCond->setPredicate(Pred);
562 LbCond->setOperand(0, TcDec);
563 LbCond->setOperand(1, cast<Value>(ConstantInt::get(Ty, 0)));
566 // Step 4: All the references to the original population counter outside
567 // the loop are replaced with the NewCount -- the value returned from
568 // __builtin_ctpop().
569 CntInst->replaceUsesOutsideBlock(NewCount, Body);
571 // step 5: Forget the "non-computable" trip-count SCEV associated with the
572 // loop. The loop would otherwise not be deleted even if it becomes empty.
573 SE->forgetLoop(CurLoop);
576 CallInst *NclPopcountRecognize::createPopcntIntrinsic(IRBuilderTy &IRBuilder,
577 Value *Val, DebugLoc DL) {
578 Value *Ops[] = { Val };
579 Type *Tys[] = { Val->getType() };
581 Module *M = (*(CurLoop->block_begin()))->getParent()->getParent();
582 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
583 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
589 /// recognize - detect population count idiom in a non-countable loop. If
590 /// detected, transform the relevant code to popcount intrinsic function
591 /// call, and return true; otherwise, return false.
592 bool NclPopcountRecognize::recognize() {
594 if (!LIR.getTargetTransformInfo())
597 LIR.getScalarEvolution();
599 if (!preliminaryScreen())
602 Instruction *CntInst;
605 if (!detectIdiom(CntInst, CntPhi, Val))
608 transform(CntInst, CntPhi, Val);
612 //===----------------------------------------------------------------------===//
614 // Implementation of LoopIdiomRecognize
616 //===----------------------------------------------------------------------===//
618 bool LoopIdiomRecognize::runOnCountableLoop() {
619 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
620 if (isa<SCEVCouldNotCompute>(BECount)) return false;
622 // If this loop executes exactly one time, then it should be peeled, not
623 // optimized by this pass.
624 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
625 if (BECst->getValue()->getValue() == 0)
628 // We require target data for now.
629 if (!getDataLayout())
633 (void)getDominatorTree();
635 LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
636 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
639 (void)getTargetLibraryInfo();
641 SmallVector<BasicBlock*, 8> ExitBlocks;
642 CurLoop->getUniqueExitBlocks(ExitBlocks);
644 DEBUG(dbgs() << "loop-idiom Scanning: F["
645 << CurLoop->getHeader()->getParent()->getName()
646 << "] Loop %" << CurLoop->getHeader()->getName() << "\n");
648 bool MadeChange = false;
649 // Scan all the blocks in the loop that are not in subloops.
650 for (Loop::block_iterator BI = CurLoop->block_begin(),
651 E = CurLoop->block_end(); BI != E; ++BI) {
652 // Ignore blocks in subloops.
653 if (LI.getLoopFor(*BI) != CurLoop)
656 MadeChange |= runOnLoopBlock(*BI, BECount, ExitBlocks);
661 bool LoopIdiomRecognize::runOnNoncountableLoop() {
662 NclPopcountRecognize Popcount(*this);
663 if (Popcount.recognize())
669 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
670 if (skipOptnoneFunction(L))
675 // If the loop could not be converted to canonical form, it must have an
676 // indirectbr in it, just give up.
677 if (!L->getLoopPreheader())
680 // Disable loop idiom recognition if the function's name is a common idiom.
681 StringRef Name = L->getHeader()->getParent()->getName();
682 if (Name == "memset" || Name == "memcpy")
685 SE = &getAnalysis<ScalarEvolution>();
686 if (SE->hasLoopInvariantBackedgeTakenCount(L))
687 return runOnCountableLoop();
688 return runOnNoncountableLoop();
691 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
692 /// with the specified backedge count. This block is known to be in the current
693 /// loop and not in any subloops.
694 bool LoopIdiomRecognize::runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
695 SmallVectorImpl<BasicBlock*> &ExitBlocks) {
696 // We can only promote stores in this block if they are unconditionally
697 // executed in the loop. For a block to be unconditionally executed, it has
698 // to dominate all the exit blocks of the loop. Verify this now.
699 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
700 if (!DT->dominates(BB, ExitBlocks[i]))
703 bool MadeChange = false;
704 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
705 Instruction *Inst = I++;
706 // Look for store instructions, which may be optimized to memset/memcpy.
707 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
709 if (!processLoopStore(SI, BECount)) continue;
712 // If processing the store invalidated our iterator, start over from the
719 // Look for memset instructions, which may be optimized to a larger memset.
720 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
722 if (!processLoopMemSet(MSI, BECount)) continue;
725 // If processing the memset invalidated our iterator, start over from the
737 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
738 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
739 if (!SI->isSimple()) return false;
741 Value *StoredVal = SI->getValueOperand();
742 Value *StorePtr = SI->getPointerOperand();
744 // Reject stores that are so large that they overflow an unsigned.
745 uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
746 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
749 // See if the pointer expression is an AddRec like {base,+,1} on the current
750 // loop, which indicates a strided store. If we have something else, it's a
751 // random store we can't handle.
752 const SCEVAddRecExpr *StoreEv =
753 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
754 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
757 // Check to see if the stride matches the size of the store. If so, then we
758 // know that every byte is touched in the loop.
759 unsigned StoreSize = (unsigned)SizeInBits >> 3;
760 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
762 if (!Stride || StoreSize != Stride->getValue()->getValue()) {
763 // TODO: Could also handle negative stride here someday, that will require
764 // the validity check in mayLoopAccessLocation to be updated though.
765 // Enable this to print exact negative strides.
766 if (0 && Stride && StoreSize == -Stride->getValue()->getValue()) {
767 dbgs() << "NEGATIVE STRIDE: " << *SI << "\n";
768 dbgs() << "BB: " << *SI->getParent();
774 // See if we can optimize just this store in isolation.
775 if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
776 StoredVal, SI, StoreEv, BECount))
779 // If the stored value is a strided load in the same loop with the same stride
780 // this this may be transformable into a memcpy. This kicks in for stuff like
781 // for (i) A[i] = B[i];
782 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
783 const SCEVAddRecExpr *LoadEv =
784 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
785 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
786 StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple())
787 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
790 //errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
795 /// processLoopMemSet - See if this memset can be promoted to a large memset.
796 bool LoopIdiomRecognize::
797 processLoopMemSet(MemSetInst *MSI, const SCEV *BECount) {
798 // We can only handle non-volatile memsets with a constant size.
799 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength())) return false;
801 // If we're not allowed to hack on memset, we fail.
802 if (!TLI->has(LibFunc::memset))
805 Value *Pointer = MSI->getDest();
807 // See if the pointer expression is an AddRec like {base,+,1} on the current
808 // loop, which indicates a strided store. If we have something else, it's a
809 // random store we can't handle.
810 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
811 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
814 // Reject memsets that are so large that they overflow an unsigned.
815 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
816 if ((SizeInBytes >> 32) != 0)
819 // Check to see if the stride matches the size of the memset. If so, then we
820 // know that every byte is touched in the loop.
821 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
823 // TODO: Could also handle negative stride here someday, that will require the
824 // validity check in mayLoopAccessLocation to be updated though.
825 if (!Stride || MSI->getLength() != Stride->getValue())
828 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
829 MSI->getAlignment(), MSI->getValue(),
834 /// mayLoopAccessLocation - Return true if the specified loop might access the
835 /// specified pointer location, which is a loop-strided access. The 'Access'
836 /// argument specifies what the verboten forms of access are (read or write).
837 static bool mayLoopAccessLocation(Value *Ptr,AliasAnalysis::ModRefResult Access,
838 Loop *L, const SCEV *BECount,
839 unsigned StoreSize, AliasAnalysis &AA,
840 Instruction *IgnoredStore) {
841 // Get the location that may be stored across the loop. Since the access is
842 // strided positively through memory, we say that the modified location starts
843 // at the pointer and has infinite size.
844 uint64_t AccessSize = AliasAnalysis::UnknownSize;
846 // If the loop iterates a fixed number of times, we can refine the access size
847 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
848 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
849 AccessSize = (BECst->getValue()->getZExtValue()+1)*StoreSize;
851 // TODO: For this to be really effective, we have to dive into the pointer
852 // operand in the store. Store to &A[i] of 100 will always return may alias
853 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
854 // which will then no-alias a store to &A[100].
855 AliasAnalysis::Location StoreLoc(Ptr, AccessSize);
857 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
859 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
860 if (&*I != IgnoredStore &&
861 (AA.getModRefInfo(I, StoreLoc) & Access))
867 /// getMemSetPatternValue - If a strided store of the specified value is safe to
868 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
869 /// be passed in. Otherwise, return null.
871 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
872 /// just replicate their input array and then pass on to memset_pattern16.
873 static Constant *getMemSetPatternValue(Value *V, const DataLayout &DL) {
874 // If the value isn't a constant, we can't promote it to being in a constant
875 // array. We could theoretically do a store to an alloca or something, but
876 // that doesn't seem worthwhile.
877 Constant *C = dyn_cast<Constant>(V);
878 if (!C) return nullptr;
880 // Only handle simple values that are a power of two bytes in size.
881 uint64_t Size = DL.getTypeSizeInBits(V->getType());
882 if (Size == 0 || (Size & 7) || (Size & (Size-1)))
885 // Don't care enough about darwin/ppc to implement this.
886 if (DL.isBigEndian())
889 // Convert to size in bytes.
892 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
893 // if the top and bottom are the same (e.g. for vectors and large integers).
894 if (Size > 16) return nullptr;
896 // If the constant is exactly 16 bytes, just use it.
897 if (Size == 16) return C;
899 // Otherwise, we'll use an array of the constants.
900 unsigned ArraySize = 16/Size;
901 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
902 return ConstantArray::get(AT, std::vector<Constant*>(ArraySize, C));
906 /// processLoopStridedStore - We see a strided store of some value. If we can
907 /// transform this into a memset or memset_pattern in the loop preheader, do so.
908 bool LoopIdiomRecognize::
909 processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
910 unsigned StoreAlignment, Value *StoredVal,
911 Instruction *TheStore, const SCEVAddRecExpr *Ev,
912 const SCEV *BECount) {
914 // If the stored value is a byte-wise value (like i32 -1), then it may be
915 // turned into a memset of i8 -1, assuming that all the consecutive bytes
916 // are stored. A store of i32 0x01020304 can never be turned into a memset,
917 // but it can be turned into memset_pattern if the target supports it.
918 Value *SplatValue = isBytewiseValue(StoredVal);
919 Constant *PatternValue = nullptr;
921 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
923 // If we're allowed to form a memset, and the stored value would be acceptable
924 // for memset, use it.
925 if (SplatValue && TLI->has(LibFunc::memset) &&
926 // Verify that the stored value is loop invariant. If not, we can't
927 // promote the memset.
928 CurLoop->isLoopInvariant(SplatValue)) {
929 // Keep and use SplatValue.
930 PatternValue = nullptr;
931 } else if (DestAS == 0 &&
932 TLI->has(LibFunc::memset_pattern16) &&
933 (PatternValue = getMemSetPatternValue(StoredVal, *DL))) {
934 // Don't create memset_pattern16s with address spaces.
935 // It looks like we can use PatternValue!
936 SplatValue = nullptr;
938 // Otherwise, this isn't an idiom we can transform. For example, we can't
939 // do anything with a 3-byte store.
943 // The trip count of the loop and the base pointer of the addrec SCEV is
944 // guaranteed to be loop invariant, which means that it should dominate the
945 // header. This allows us to insert code for it in the preheader.
946 BasicBlock *Preheader = CurLoop->getLoopPreheader();
947 IRBuilder<> Builder(Preheader->getTerminator());
948 SCEVExpander Expander(*SE, "loop-idiom");
950 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
952 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
953 // this into a memset in the loop preheader now if we want. However, this
954 // would be unsafe to do if there is anything else in the loop that may read
955 // or write to the aliased location. Check for any overlap by generating the
956 // base pointer and checking the region.
958 Expander.expandCodeFor(Ev->getStart(), DestInt8PtrTy,
959 Preheader->getTerminator());
961 if (mayLoopAccessLocation(BasePtr, AliasAnalysis::ModRef,
963 StoreSize, getAnalysis<AliasAnalysis>(), TheStore)) {
965 // If we generated new code for the base pointer, clean up.
966 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
970 // Okay, everything looks good, insert the memset.
972 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
973 // pointer size if it isn't already.
974 Type *IntPtr = Builder.getIntPtrTy(DL, DestAS);
975 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
977 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
979 if (StoreSize != 1) {
980 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
985 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
989 NewCall = Builder.CreateMemSet(BasePtr,
994 // Everything is emitted in default address space
995 Type *Int8PtrTy = DestInt8PtrTy;
997 Module *M = TheStore->getParent()->getParent()->getParent();
998 Value *MSP = M->getOrInsertFunction("memset_pattern16",
1005 // Otherwise we should form a memset_pattern16. PatternValue is known to be
1006 // an constant array of 16-bytes. Plop the value into a mergable global.
1007 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
1008 GlobalValue::InternalLinkage,
1009 PatternValue, ".memset_pattern");
1010 GV->setUnnamedAddr(true); // Ok to merge these.
1011 GV->setAlignment(16);
1012 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
1013 NewCall = Builder.CreateCall3(MSP, BasePtr, PatternPtr, NumBytes);
1016 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
1017 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
1018 NewCall->setDebugLoc(TheStore->getDebugLoc());
1020 // Okay, the memset has been formed. Zap the original store and anything that
1022 deleteDeadInstruction(TheStore, TLI);
1027 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
1028 /// same-strided load.
1029 bool LoopIdiomRecognize::
1030 processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
1031 const SCEVAddRecExpr *StoreEv,
1032 const SCEVAddRecExpr *LoadEv,
1033 const SCEV *BECount) {
1034 // If we're not allowed to form memcpy, we fail.
1035 if (!TLI->has(LibFunc::memcpy))
1038 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
1040 // The trip count of the loop and the base pointer of the addrec SCEV is
1041 // guaranteed to be loop invariant, which means that it should dominate the
1042 // header. This allows us to insert code for it in the preheader.
1043 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1044 IRBuilder<> Builder(Preheader->getTerminator());
1045 SCEVExpander Expander(*SE, "loop-idiom");
1047 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
1048 // this into a memcpy in the loop preheader now if we want. However, this
1049 // would be unsafe to do if there is anything else in the loop that may read
1050 // or write the memory region we're storing to. This includes the load that
1051 // feeds the stores. Check for an alias by generating the base address and
1052 // checking everything.
1053 Value *StoreBasePtr =
1054 Expander.expandCodeFor(StoreEv->getStart(),
1055 Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
1056 Preheader->getTerminator());
1058 if (mayLoopAccessLocation(StoreBasePtr, AliasAnalysis::ModRef,
1059 CurLoop, BECount, StoreSize,
1060 getAnalysis<AliasAnalysis>(), SI)) {
1062 // If we generated new code for the base pointer, clean up.
1063 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
1067 // For a memcpy, we have to make sure that the input array is not being
1068 // mutated by the loop.
1069 Value *LoadBasePtr =
1070 Expander.expandCodeFor(LoadEv->getStart(),
1071 Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
1072 Preheader->getTerminator());
1074 if (mayLoopAccessLocation(LoadBasePtr, AliasAnalysis::Mod, CurLoop, BECount,
1075 StoreSize, getAnalysis<AliasAnalysis>(), SI)) {
1077 // If we generated new code for the base pointer, clean up.
1078 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
1079 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
1083 // Okay, everything is safe, we can transform this!
1086 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
1087 // pointer size if it isn't already.
1088 Type *IntPtrTy = Builder.getIntPtrTy(DL, SI->getPointerAddressSpace());
1089 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
1091 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtrTy, 1),
1094 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
1098 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
1101 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
1102 std::min(SI->getAlignment(), LI->getAlignment()));
1103 NewCall->setDebugLoc(SI->getDebugLoc());
1105 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
1106 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
1107 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
1110 // Okay, the memset has been formed. Zap the original store and anything that
1112 deleteDeadInstruction(SI, TLI);