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/TargetLibraryInfo.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/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 {
135 TargetLibraryInfo *TLI;
136 const TargetTransformInfo *TTI;
139 explicit LoopIdiomRecognize() : LoopPass(ID) {
140 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
147 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
148 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
149 SmallVectorImpl<BasicBlock*> &ExitBlocks);
151 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
152 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
154 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
155 unsigned StoreAlignment,
156 Value *SplatValue, Instruction *TheStore,
157 const SCEVAddRecExpr *Ev,
158 const SCEV *BECount);
159 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
160 const SCEVAddRecExpr *StoreEv,
161 const SCEVAddRecExpr *LoadEv,
162 const SCEV *BECount);
164 /// This transformation requires natural loop information & requires that
165 /// loop preheaders be inserted into the CFG.
167 void getAnalysisUsage(AnalysisUsage &AU) const override {
168 AU.addRequired<LoopInfoWrapperPass>();
169 AU.addPreserved<LoopInfoWrapperPass>();
170 AU.addRequiredID(LoopSimplifyID);
171 AU.addPreservedID(LoopSimplifyID);
172 AU.addRequiredID(LCSSAID);
173 AU.addPreservedID(LCSSAID);
174 AU.addRequired<AliasAnalysis>();
175 AU.addPreserved<AliasAnalysis>();
176 AU.addRequired<ScalarEvolution>();
177 AU.addPreserved<ScalarEvolution>();
178 AU.addPreserved<DominatorTreeWrapperPass>();
179 AU.addRequired<DominatorTreeWrapperPass>();
180 AU.addRequired<TargetLibraryInfoWrapperPass>();
181 AU.addRequired<TargetTransformInfoWrapperPass>();
184 DominatorTree *getDominatorTree() {
186 : (DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree());
189 ScalarEvolution *getScalarEvolution() {
190 return SE ? SE : (SE = &getAnalysis<ScalarEvolution>());
193 TargetLibraryInfo *getTargetLibraryInfo() {
195 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
200 const TargetTransformInfo *getTargetTransformInfo() {
202 : (TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
203 *CurLoop->getHeader()->getParent()));
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(LoopInfoWrapperPass)
218 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
219 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
220 INITIALIZE_PASS_DEPENDENCY(LCSSA)
221 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
222 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
223 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
224 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
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,
235 const TargetLibraryInfo *TLI) {
236 SmallVector<Value *, 16> Operands(I->value_op_begin(), I->value_op_end());
237 I->replaceAllUsesWith(UndefValue::get(I->getType()));
238 I->eraseFromParent();
239 for (Value *Op : Operands)
240 RecursivelyDeleteTriviallyDeadInstructions(Op, TLI);
243 //===----------------------------------------------------------------------===//
245 // Implementation of LIRUtil
247 //===----------------------------------------------------------------------===//
249 // This function will return true iff the given block contains nothing but goto.
250 // A typical usage of this function is to check if the preheader function is
251 // "almost" empty such that generated intrinsic functions can be moved across
252 // the preheader and be placed at the end of the precondition block without
253 // the concern of breaking data dependence.
254 bool LIRUtil::isAlmostEmpty(BasicBlock *BB) {
255 if (BranchInst *Br = getBranch(BB)) {
256 return Br->isUnconditional() && Br == BB->begin();
261 BasicBlock *LIRUtil::getPrecondBb(BasicBlock *PreHead) {
262 if (BasicBlock *BB = PreHead->getSinglePredecessor()) {
263 BranchInst *Br = getBranch(BB);
264 return Br && Br->isConditional() ? BB : nullptr;
269 //===----------------------------------------------------------------------===//
271 // Implementation of NclPopcountRecognize
273 //===----------------------------------------------------------------------===//
275 NclPopcountRecognize::NclPopcountRecognize(LoopIdiomRecognize &TheLIR):
276 LIR(TheLIR), CurLoop(TheLIR.getLoop()), PreCondBB(nullptr) {
279 bool NclPopcountRecognize::preliminaryScreen() {
280 const TargetTransformInfo *TTI = LIR.getTargetTransformInfo();
281 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
284 // Counting population are usually conducted by few arithmetic instructions.
285 // Such instructions can be easilly "absorbed" by vacant slots in a
286 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
287 // in a compact loop.
289 // Give up if the loop has multiple blocks or multiple backedges.
290 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
293 BasicBlock *LoopBody = *(CurLoop->block_begin());
294 if (LoopBody->size() >= 20) {
295 // The loop is too big, bail out.
299 // It should have a preheader containing nothing but a goto instruction.
300 BasicBlock *PreHead = CurLoop->getLoopPreheader();
301 if (!PreHead || !LIRUtil::isAlmostEmpty(PreHead))
304 // It should have a precondition block where the generated popcount instrinsic
305 // function will be inserted.
306 PreCondBB = LIRUtil::getPrecondBb(PreHead);
313 Value *NclPopcountRecognize::matchCondition(BranchInst *Br,
314 BasicBlock *LoopEntry) const {
315 if (!Br || !Br->isConditional())
318 ICmpInst *Cond = dyn_cast<ICmpInst>(Br->getCondition());
322 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
323 if (!CmpZero || !CmpZero->isZero())
326 ICmpInst::Predicate Pred = Cond->getPredicate();
327 if ((Pred == ICmpInst::ICMP_NE && Br->getSuccessor(0) == LoopEntry) ||
328 (Pred == ICmpInst::ICMP_EQ && Br->getSuccessor(1) == LoopEntry))
329 return Cond->getOperand(0);
334 bool NclPopcountRecognize::detectIdiom(Instruction *&CntInst,
337 // Following code tries to detect this idiom:
340 // goto loop-exit // the precondition of the loop
343 // x1 = phi (x0, x2);
344 // cnt1 = phi(cnt0, cnt2);
348 // x2 = x1 & (x1 - 1);
355 // step 1: Check to see if the look-back branch match this pattern:
356 // "if (a!=0) goto loop-entry".
357 BasicBlock *LoopEntry;
358 Instruction *DefX2, *CountInst;
359 Value *VarX1, *VarX0;
360 PHINode *PhiX, *CountPhi;
362 DefX2 = CountInst = nullptr;
363 VarX1 = VarX0 = nullptr;
364 PhiX = CountPhi = nullptr;
365 LoopEntry = *(CurLoop->block_begin());
367 // step 1: Check if the loop-back branch is in desirable form.
369 if (Value *T = matchCondition (LIRUtil::getBranch(LoopEntry), LoopEntry))
370 DefX2 = dyn_cast<Instruction>(T);
375 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
377 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
380 BinaryOperator *SubOneOp;
382 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
383 VarX1 = DefX2->getOperand(1);
385 VarX1 = DefX2->getOperand(0);
386 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
391 Instruction *SubInst = cast<Instruction>(SubOneOp);
392 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
394 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
395 (SubInst->getOpcode() == Instruction::Add && Dec->isAllOnesValue()))) {
400 // step 3: Check the recurrence of variable X
402 PhiX = dyn_cast<PHINode>(VarX1);
404 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
409 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
412 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI(),
413 IterE = LoopEntry->end(); Iter != IterE; Iter++) {
414 Instruction *Inst = Iter;
415 if (Inst->getOpcode() != Instruction::Add)
418 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
419 if (!Inc || !Inc->isOne())
422 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
423 if (!Phi || Phi->getParent() != LoopEntry)
426 // Check if the result of the instruction is live of the loop.
427 bool LiveOutLoop = false;
428 for (User *U : Inst->users()) {
429 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
430 LiveOutLoop = true; break;
445 // step 5: check if the precondition is in this form:
446 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
448 BranchInst *PreCondBr = LIRUtil::getBranch(PreCondBB);
449 Value *T = matchCondition (PreCondBr, CurLoop->getLoopPreheader());
450 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
461 void NclPopcountRecognize::transform(Instruction *CntInst,
462 PHINode *CntPhi, Value *Var) {
464 ScalarEvolution *SE = LIR.getScalarEvolution();
465 TargetLibraryInfo *TLI = LIR.getTargetLibraryInfo();
466 BasicBlock *PreHead = CurLoop->getLoopPreheader();
467 BranchInst *PreCondBr = LIRUtil::getBranch(PreCondBB);
468 const DebugLoc DL = CntInst->getDebugLoc();
470 // Assuming before transformation, the loop is following:
471 // if (x) // the precondition
472 // do { cnt++; x &= x - 1; } while(x);
474 // Step 1: Insert the ctpop instruction at the end of the precondition block
475 IRBuilderTy Builder(PreCondBr);
476 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
478 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
479 NewCount = PopCntZext =
480 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
482 if (NewCount != PopCnt)
483 (cast<Instruction>(NewCount))->setDebugLoc(DL);
485 // TripCnt is exactly the number of iterations the loop has
488 // If the population counter's initial value is not zero, insert Add Inst.
489 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
490 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
491 if (!InitConst || !InitConst->isZero()) {
492 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
493 (cast<Instruction>(NewCount))->setDebugLoc(DL);
497 // Step 2: Replace the precondition from "if(x == 0) goto loop-exit" to
498 // "if(NewCount == 0) loop-exit". Withtout this change, the intrinsic
499 // function would be partial dead code, and downstream passes will drag
500 // it back from the precondition block to the preheader.
502 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
504 Value *Opnd0 = PopCntZext;
505 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
506 if (PreCond->getOperand(0) != Var)
507 std::swap(Opnd0, Opnd1);
509 ICmpInst *NewPreCond =
510 cast<ICmpInst>(Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
511 PreCond->replaceAllUsesWith(NewPreCond);
513 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
516 // Step 3: Note that the population count is exactly the trip count of the
517 // loop in question, which enble us to to convert the loop from noncountable
518 // loop into a countable one. The benefit is twofold:
520 // - If the loop only counts population, the entire loop become dead after
521 // the transformation. It is lots easier to prove a countable loop dead
522 // than to prove a noncountable one. (In some C dialects, a infite loop
523 // isn't dead even if it computes nothing useful. In general, DCE needs
524 // to prove a noncountable loop finite before safely delete it.)
526 // - If the loop also performs something else, it remains alive.
527 // Since it is transformed to countable form, it can be aggressively
528 // optimized by some optimizations which are in general not applicable
529 // to a noncountable loop.
531 // After this step, this loop (conceptually) would look like following:
532 // newcnt = __builtin_ctpop(x);
535 // do { cnt++; x &= x-1; t--) } while (t > 0);
536 BasicBlock *Body = *(CurLoop->block_begin());
538 BranchInst *LbBr = LIRUtil::getBranch(Body);
539 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
540 Type *Ty = TripCnt->getType();
542 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", Body->begin());
544 Builder.SetInsertPoint(LbCond);
545 Value *Opnd1 = cast<Value>(TcPhi);
546 Value *Opnd2 = cast<Value>(ConstantInt::get(Ty, 1));
548 cast<Instruction>(Builder.CreateSub(Opnd1, Opnd2, "tcdec", false, true));
550 TcPhi->addIncoming(TripCnt, PreHead);
551 TcPhi->addIncoming(TcDec, Body);
553 CmpInst::Predicate Pred = (LbBr->getSuccessor(0) == Body) ?
554 CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
555 LbCond->setPredicate(Pred);
556 LbCond->setOperand(0, TcDec);
557 LbCond->setOperand(1, cast<Value>(ConstantInt::get(Ty, 0)));
560 // Step 4: All the references to the original population counter outside
561 // the loop are replaced with the NewCount -- the value returned from
562 // __builtin_ctpop().
563 CntInst->replaceUsesOutsideBlock(NewCount, Body);
565 // step 5: Forget the "non-computable" trip-count SCEV associated with the
566 // loop. The loop would otherwise not be deleted even if it becomes empty.
567 SE->forgetLoop(CurLoop);
570 CallInst *NclPopcountRecognize::createPopcntIntrinsic(IRBuilderTy &IRBuilder,
571 Value *Val, DebugLoc DL) {
572 Value *Ops[] = { Val };
573 Type *Tys[] = { Val->getType() };
575 Module *M = (*(CurLoop->block_begin()))->getParent()->getParent();
576 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
577 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
583 /// recognize - detect population count idiom in a non-countable loop. If
584 /// detected, transform the relevant code to popcount intrinsic function
585 /// call, and return true; otherwise, return false.
586 bool NclPopcountRecognize::recognize() {
588 if (!LIR.getTargetTransformInfo())
591 LIR.getScalarEvolution();
593 if (!preliminaryScreen())
596 Instruction *CntInst;
599 if (!detectIdiom(CntInst, CntPhi, Val))
602 transform(CntInst, CntPhi, Val);
606 //===----------------------------------------------------------------------===//
608 // Implementation of LoopIdiomRecognize
610 //===----------------------------------------------------------------------===//
612 bool LoopIdiomRecognize::runOnCountableLoop() {
613 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
614 assert(!isa<SCEVCouldNotCompute>(BECount) &&
615 "runOnCountableLoop() called on a loop without a predictable"
616 "backedge-taken count");
618 // If this loop executes exactly one time, then it should be peeled, not
619 // optimized by this pass.
620 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
621 if (BECst->getValue()->getValue() == 0)
625 (void)getDominatorTree();
627 LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
628 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
631 (void)getTargetLibraryInfo();
633 SmallVector<BasicBlock*, 8> ExitBlocks;
634 CurLoop->getUniqueExitBlocks(ExitBlocks);
636 DEBUG(dbgs() << "loop-idiom Scanning: F["
637 << CurLoop->getHeader()->getParent()->getName()
638 << "] Loop %" << CurLoop->getHeader()->getName() << "\n");
640 bool MadeChange = false;
641 // Scan all the blocks in the loop that are not in subloops.
642 for (auto *BB : CurLoop->getBlocks()) {
643 // Ignore blocks in subloops.
644 if (LI.getLoopFor(BB) != CurLoop)
647 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
652 bool LoopIdiomRecognize::runOnNoncountableLoop() {
653 NclPopcountRecognize Popcount(*this);
654 if (Popcount.recognize())
660 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
661 if (skipOptnoneFunction(L))
666 // If the loop could not be converted to canonical form, it must have an
667 // indirectbr in it, just give up.
668 if (!L->getLoopPreheader())
671 // Disable loop idiom recognition if the function's name is a common idiom.
672 StringRef Name = L->getHeader()->getParent()->getName();
673 if (Name == "memset" || Name == "memcpy")
676 SE = &getAnalysis<ScalarEvolution>();
677 if (SE->hasLoopInvariantBackedgeTakenCount(L))
678 return runOnCountableLoop();
679 return runOnNoncountableLoop();
682 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
683 /// with the specified backedge count. This block is known to be in the current
684 /// loop and not in any subloops.
685 bool LoopIdiomRecognize::runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
686 SmallVectorImpl<BasicBlock*> &ExitBlocks) {
687 // We can only promote stores in this block if they are unconditionally
688 // executed in the loop. For a block to be unconditionally executed, it has
689 // to dominate all the exit blocks of the loop. Verify this now.
690 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
691 if (!DT->dominates(BB, ExitBlocks[i]))
694 bool MadeChange = false;
695 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
696 Instruction *Inst = I++;
697 // Look for store instructions, which may be optimized to memset/memcpy.
698 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
700 if (!processLoopStore(SI, BECount)) continue;
703 // If processing the store invalidated our iterator, start over from the
710 // Look for memset instructions, which may be optimized to a larger memset.
711 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
713 if (!processLoopMemSet(MSI, BECount)) continue;
716 // If processing the memset invalidated our iterator, start over from the
728 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
729 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
730 if (!SI->isSimple()) return false;
732 Value *StoredVal = SI->getValueOperand();
733 Value *StorePtr = SI->getPointerOperand();
735 // Reject stores that are so large that they overflow an unsigned.
736 auto &DL = CurLoop->getHeader()->getModule()->getDataLayout();
737 uint64_t SizeInBits = DL.getTypeSizeInBits(StoredVal->getType());
738 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
741 // See if the pointer expression is an AddRec like {base,+,1} on the current
742 // loop, which indicates a strided store. If we have something else, it's a
743 // random store we can't handle.
744 const SCEVAddRecExpr *StoreEv =
745 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
746 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
749 // Check to see if the stride matches the size of the store. If so, then we
750 // know that every byte is touched in the loop.
751 unsigned StoreSize = (unsigned)SizeInBits >> 3;
752 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
754 if (!Stride || StoreSize != Stride->getValue()->getValue()) {
755 // TODO: Could also handle negative stride here someday, that will require
756 // the validity check in mayLoopAccessLocation to be updated though.
757 // Enable this to print exact negative strides.
758 if (0 && Stride && StoreSize == -Stride->getValue()->getValue()) {
759 dbgs() << "NEGATIVE STRIDE: " << *SI << "\n";
760 dbgs() << "BB: " << *SI->getParent();
766 // See if we can optimize just this store in isolation.
767 if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
768 StoredVal, SI, StoreEv, BECount))
771 // If the stored value is a strided load in the same loop with the same stride
772 // this this may be transformable into a memcpy. This kicks in for stuff like
773 // for (i) A[i] = B[i];
774 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
775 const SCEVAddRecExpr *LoadEv =
776 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
777 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
778 StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple())
779 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
782 //errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
787 /// processLoopMemSet - See if this memset can be promoted to a large memset.
788 bool LoopIdiomRecognize::
789 processLoopMemSet(MemSetInst *MSI, const SCEV *BECount) {
790 // We can only handle non-volatile memsets with a constant size.
791 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength())) return false;
793 // If we're not allowed to hack on memset, we fail.
794 if (!TLI->has(LibFunc::memset))
797 Value *Pointer = MSI->getDest();
799 // See if the pointer expression is an AddRec like {base,+,1} on the current
800 // loop, which indicates a strided store. If we have something else, it's a
801 // random store we can't handle.
802 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
803 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
806 // Reject memsets that are so large that they overflow an unsigned.
807 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
808 if ((SizeInBytes >> 32) != 0)
811 // Check to see if the stride matches the size of the memset. If so, then we
812 // know that every byte is touched in the loop.
813 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
815 // TODO: Could also handle negative stride here someday, that will require the
816 // validity check in mayLoopAccessLocation to be updated though.
817 if (!Stride || MSI->getLength() != Stride->getValue())
820 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
821 MSI->getAlignment(), MSI->getValue(),
826 /// mayLoopAccessLocation - Return true if the specified loop might access the
827 /// specified pointer location, which is a loop-strided access. The 'Access'
828 /// argument specifies what the verboten forms of access are (read or write).
829 static bool mayLoopAccessLocation(Value *Ptr,AliasAnalysis::ModRefResult Access,
830 Loop *L, const SCEV *BECount,
831 unsigned StoreSize, AliasAnalysis &AA,
832 Instruction *IgnoredStore) {
833 // Get the location that may be stored across the loop. Since the access is
834 // strided positively through memory, we say that the modified location starts
835 // at the pointer and has infinite size.
836 uint64_t AccessSize = MemoryLocation::UnknownSize;
838 // If the loop iterates a fixed number of times, we can refine the access size
839 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
840 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
841 AccessSize = (BECst->getValue()->getZExtValue()+1)*StoreSize;
843 // TODO: For this to be really effective, we have to dive into the pointer
844 // operand in the store. Store to &A[i] of 100 will always return may alias
845 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
846 // which will then no-alias a store to &A[100].
847 MemoryLocation StoreLoc(Ptr, AccessSize);
849 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
851 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
852 if (&*I != IgnoredStore &&
853 (AA.getModRefInfo(I, StoreLoc) & Access))
859 /// getMemSetPatternValue - If a strided store of the specified value is safe to
860 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
861 /// be passed in. Otherwise, return null.
863 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
864 /// just replicate their input array and then pass on to memset_pattern16.
865 static Constant *getMemSetPatternValue(Value *V, const DataLayout &DL) {
866 // If the value isn't a constant, we can't promote it to being in a constant
867 // array. We could theoretically do a store to an alloca or something, but
868 // that doesn't seem worthwhile.
869 Constant *C = dyn_cast<Constant>(V);
870 if (!C) return nullptr;
872 // Only handle simple values that are a power of two bytes in size.
873 uint64_t Size = DL.getTypeSizeInBits(V->getType());
874 if (Size == 0 || (Size & 7) || (Size & (Size-1)))
877 // Don't care enough about darwin/ppc to implement this.
878 if (DL.isBigEndian())
881 // Convert to size in bytes.
884 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
885 // if the top and bottom are the same (e.g. for vectors and large integers).
886 if (Size > 16) return nullptr;
888 // If the constant is exactly 16 bytes, just use it.
889 if (Size == 16) return C;
891 // Otherwise, we'll use an array of the constants.
892 unsigned ArraySize = 16/Size;
893 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
894 return ConstantArray::get(AT, std::vector<Constant*>(ArraySize, C));
898 /// processLoopStridedStore - We see a strided store of some value. If we can
899 /// transform this into a memset or memset_pattern in the loop preheader, do so.
900 bool LoopIdiomRecognize::
901 processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
902 unsigned StoreAlignment, Value *StoredVal,
903 Instruction *TheStore, const SCEVAddRecExpr *Ev,
904 const SCEV *BECount) {
906 // If the stored value is a byte-wise value (like i32 -1), then it may be
907 // turned into a memset of i8 -1, assuming that all the consecutive bytes
908 // are stored. A store of i32 0x01020304 can never be turned into a memset,
909 // but it can be turned into memset_pattern if the target supports it.
910 Value *SplatValue = isBytewiseValue(StoredVal);
911 Constant *PatternValue = nullptr;
912 auto &DL = CurLoop->getHeader()->getModule()->getDataLayout();
913 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
915 // If we're allowed to form a memset, and the stored value would be acceptable
916 // for memset, use it.
917 if (SplatValue && TLI->has(LibFunc::memset) &&
918 // Verify that the stored value is loop invariant. If not, we can't
919 // promote the memset.
920 CurLoop->isLoopInvariant(SplatValue)) {
921 // Keep and use SplatValue.
922 PatternValue = nullptr;
923 } else if (DestAS == 0 && TLI->has(LibFunc::memset_pattern16) &&
924 (PatternValue = getMemSetPatternValue(StoredVal, DL))) {
925 // Don't create memset_pattern16s with address spaces.
926 // It looks like we can use PatternValue!
927 SplatValue = nullptr;
929 // Otherwise, this isn't an idiom we can transform. For example, we can't
930 // do anything with a 3-byte store.
934 // The trip count of the loop and the base pointer of the addrec SCEV is
935 // guaranteed to be loop invariant, which means that it should dominate the
936 // header. This allows us to insert code for it in the preheader.
937 BasicBlock *Preheader = CurLoop->getLoopPreheader();
938 IRBuilder<> Builder(Preheader->getTerminator());
939 SCEVExpander Expander(*SE, DL, "loop-idiom");
941 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
943 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
944 // this into a memset in the loop preheader now if we want. However, this
945 // would be unsafe to do if there is anything else in the loop that may read
946 // or write to the aliased location. Check for any overlap by generating the
947 // base pointer and checking the region.
949 Expander.expandCodeFor(Ev->getStart(), DestInt8PtrTy,
950 Preheader->getTerminator());
952 if (mayLoopAccessLocation(BasePtr, AliasAnalysis::ModRef,
954 StoreSize, getAnalysis<AliasAnalysis>(), TheStore)) {
956 // If we generated new code for the base pointer, clean up.
957 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
961 // Okay, everything looks good, insert the memset.
963 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
964 // pointer size if it isn't already.
965 Type *IntPtr = Builder.getIntPtrTy(DL, DestAS);
966 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
968 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
970 if (StoreSize != 1) {
971 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
976 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
980 NewCall = Builder.CreateMemSet(BasePtr,
985 // Everything is emitted in default address space
986 Type *Int8PtrTy = DestInt8PtrTy;
988 Module *M = TheStore->getParent()->getParent()->getParent();
989 Value *MSP = M->getOrInsertFunction("memset_pattern16",
996 // Otherwise we should form a memset_pattern16. PatternValue is known to be
997 // an constant array of 16-bytes. Plop the value into a mergable global.
998 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
999 GlobalValue::PrivateLinkage,
1000 PatternValue, ".memset_pattern");
1001 GV->setUnnamedAddr(true); // Ok to merge these.
1002 GV->setAlignment(16);
1003 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
1004 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
1007 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
1008 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
1009 NewCall->setDebugLoc(TheStore->getDebugLoc());
1011 // Okay, the memset has been formed. Zap the original store and anything that
1013 deleteDeadInstruction(TheStore, TLI);
1018 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
1019 /// same-strided load.
1020 bool LoopIdiomRecognize::
1021 processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
1022 const SCEVAddRecExpr *StoreEv,
1023 const SCEVAddRecExpr *LoadEv,
1024 const SCEV *BECount) {
1025 // If we're not allowed to form memcpy, we fail.
1026 if (!TLI->has(LibFunc::memcpy))
1029 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
1031 // The trip count of the loop and the base pointer of the addrec SCEV is
1032 // guaranteed to be loop invariant, which means that it should dominate the
1033 // header. This allows us to insert code for it in the preheader.
1034 BasicBlock *Preheader = CurLoop->getLoopPreheader();
1035 IRBuilder<> Builder(Preheader->getTerminator());
1036 const DataLayout &DL = Preheader->getModule()->getDataLayout();
1037 SCEVExpander Expander(*SE, DL, "loop-idiom");
1039 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
1040 // this into a memcpy in the loop preheader now if we want. However, this
1041 // would be unsafe to do if there is anything else in the loop that may read
1042 // or write the memory region we're storing to. This includes the load that
1043 // feeds the stores. Check for an alias by generating the base address and
1044 // checking everything.
1045 Value *StoreBasePtr =
1046 Expander.expandCodeFor(StoreEv->getStart(),
1047 Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
1048 Preheader->getTerminator());
1050 if (mayLoopAccessLocation(StoreBasePtr, AliasAnalysis::ModRef,
1051 CurLoop, BECount, StoreSize,
1052 getAnalysis<AliasAnalysis>(), SI)) {
1054 // If we generated new code for the base pointer, clean up.
1055 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
1059 // For a memcpy, we have to make sure that the input array is not being
1060 // mutated by the loop.
1061 Value *LoadBasePtr =
1062 Expander.expandCodeFor(LoadEv->getStart(),
1063 Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
1064 Preheader->getTerminator());
1066 if (mayLoopAccessLocation(LoadBasePtr, AliasAnalysis::Mod, CurLoop, BECount,
1067 StoreSize, getAnalysis<AliasAnalysis>(), SI)) {
1069 // If we generated new code for the base pointer, clean up.
1070 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
1071 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
1075 // Okay, everything is safe, we can transform this!
1078 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
1079 // pointer size if it isn't already.
1080 Type *IntPtrTy = Builder.getIntPtrTy(DL, SI->getPointerAddressSpace());
1081 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
1083 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtrTy, 1),
1086 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
1090 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
1093 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
1094 std::min(SI->getAlignment(), LI->getAlignment()));
1095 NewCall->setDebugLoc(SI->getDebugLoc());
1097 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
1098 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
1099 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
1102 // Okay, the memset has been formed. Zap the original store and anything that
1104 deleteDeadInstruction(SI, TLI);