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 is to recoginize idioms of population-count conducted in
73 /// a noncountable loop. Currently it only recognizes this pattern:
75 /// while(x) {cnt++; ...; x &= x - 1; ...}
77 class NclPopcountRecognize {
78 LoopIdiomRecognize &LIR;
80 BasicBlock *PreCondBB;
82 typedef IRBuilder<> IRBuilderTy;
85 explicit NclPopcountRecognize(LoopIdiomRecognize &TheLIR);
89 /// Take a glimpse of the loop to see if we need to go ahead recoginizing
91 bool preliminaryScreen();
93 /// Check if the given conditional branch is based on the comparison
94 /// between a variable and zero, and if the variable is non-zero, the
95 /// control yields to the loop entry. If the branch matches the behavior,
96 /// the variable involved in the comparion is returned. This function will
97 /// be called to see if the precondition and postcondition of the loop
98 /// are in desirable form.
99 Value *matchCondition(BranchInst *Br, BasicBlock *NonZeroTarget) const;
101 /// Return true iff the idiom is detected in the loop. and 1) \p CntInst
102 /// is set to the instruction counting the population bit. 2) \p CntPhi
103 /// is set to the corresponding phi node. 3) \p Var is set to the value
104 /// whose population bits are being counted.
105 bool detectIdiom(Instruction *&CntInst, PHINode *&CntPhi, Value *&Var) const;
107 /// Insert ctpop intrinsic function and some obviously dead instructions.
108 void transform(Instruction *CntInst, PHINode *CntPhi, Value *Var);
110 /// Create llvm.ctpop.* intrinsic function.
111 CallInst *createPopcntIntrinsic(IRBuilderTy &IRB, Value *Val, DebugLoc DL);
114 class LoopIdiomRecognize : public LoopPass {
118 TargetLibraryInfo *TLI;
119 const TargetTransformInfo *TTI;
123 explicit LoopIdiomRecognize() : LoopPass(ID) {
124 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
131 bool runOnLoop(Loop *L, LPPassManager &LPM) override;
133 /// This transformation requires natural loop information & requires that
134 /// loop preheaders be inserted into the CFG.
136 void getAnalysisUsage(AnalysisUsage &AU) const override {
137 AU.addRequired<LoopInfoWrapperPass>();
138 AU.addPreserved<LoopInfoWrapperPass>();
139 AU.addRequiredID(LoopSimplifyID);
140 AU.addPreservedID(LoopSimplifyID);
141 AU.addRequiredID(LCSSAID);
142 AU.addPreservedID(LCSSAID);
143 AU.addRequired<AliasAnalysis>();
144 AU.addPreserved<AliasAnalysis>();
145 AU.addRequired<ScalarEvolution>();
146 AU.addPreserved<ScalarEvolution>();
147 AU.addPreserved<DominatorTreeWrapperPass>();
148 AU.addRequired<DominatorTreeWrapperPass>();
149 AU.addRequired<TargetLibraryInfoWrapperPass>();
150 AU.addRequired<TargetTransformInfoWrapperPass>();
153 DominatorTree *getDominatorTree() {
155 : (DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree());
158 ScalarEvolution *getScalarEvolution() {
159 return SE ? SE : (SE = &getAnalysis<ScalarEvolution>());
162 TargetLibraryInfo *getTargetLibraryInfo() {
164 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
169 const TargetTransformInfo *getTargetTransformInfo() {
171 : (TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
172 *CurLoop->getHeader()->getParent()));
175 Loop *getLoop() const { return CurLoop; }
178 /// \name Countable Loop Idiom Handling
181 bool runOnCountableLoop();
182 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
183 SmallVectorImpl<BasicBlock *> &ExitBlocks);
185 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
186 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
188 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
189 unsigned StoreAlignment, Value *SplatValue,
190 Instruction *TheStore, const SCEVAddRecExpr *Ev,
191 const SCEV *BECount);
192 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
193 const SCEVAddRecExpr *StoreEv,
194 const SCEVAddRecExpr *LoadEv,
195 const SCEV *BECount);
198 /// \name Noncountable Loop Idiom Handling
201 bool runOnNoncountableLoop();
206 } // End anonymous namespace.
208 char LoopIdiomRecognize::ID = 0;
209 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
211 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
212 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
213 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
214 INITIALIZE_PASS_DEPENDENCY(LCSSA)
215 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
216 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
217 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
218 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
219 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
222 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
224 /// deleteDeadInstruction - Delete this instruction. Before we do, go through
225 /// and zero out all the operands of this instruction. If any of them become
226 /// dead, delete them and the computation tree that feeds them.
228 static void deleteDeadInstruction(Instruction *I,
229 const TargetLibraryInfo *TLI) {
230 SmallVector<Value *, 16> Operands(I->value_op_begin(), I->value_op_end());
231 I->replaceAllUsesWith(UndefValue::get(I->getType()));
232 I->eraseFromParent();
233 for (Value *Op : Operands)
234 RecursivelyDeleteTriviallyDeadInstructions(Op, TLI);
237 //===----------------------------------------------------------------------===//
239 // Implementation of NclPopcountRecognize
241 //===----------------------------------------------------------------------===//
243 NclPopcountRecognize::NclPopcountRecognize(LoopIdiomRecognize &TheLIR)
244 : LIR(TheLIR), CurLoop(TheLIR.getLoop()), PreCondBB(nullptr) {}
246 bool NclPopcountRecognize::preliminaryScreen() {
247 const TargetTransformInfo *TTI = LIR.getTargetTransformInfo();
248 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
251 // Counting population are usually conducted by few arithmetic instructions.
252 // Such instructions can be easilly "absorbed" by vacant slots in a
253 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
254 // in a compact loop.
256 // Give up if the loop has multiple blocks or multiple backedges.
257 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
260 BasicBlock *LoopBody = *(CurLoop->block_begin());
261 if (LoopBody->size() >= 20) {
262 // The loop is too big, bail out.
266 // It should have a preheader containing nothing but an unconditional branch.
267 BasicBlock *PH = CurLoop->getLoopPreheader();
270 if (&PH->front() != PH->getTerminator())
272 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
273 if (!EntryBI || EntryBI->isConditional())
276 // It should have a precondition block where the generated popcount instrinsic
277 // function can be inserted.
278 PreCondBB = PH->getSinglePredecessor();
281 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
282 if (!PreCondBI || PreCondBI->isUnconditional())
288 Value *NclPopcountRecognize::matchCondition(BranchInst *Br,
289 BasicBlock *LoopEntry) const {
290 if (!Br || !Br->isConditional())
293 ICmpInst *Cond = dyn_cast<ICmpInst>(Br->getCondition());
297 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
298 if (!CmpZero || !CmpZero->isZero())
301 ICmpInst::Predicate Pred = Cond->getPredicate();
302 if ((Pred == ICmpInst::ICMP_NE && Br->getSuccessor(0) == LoopEntry) ||
303 (Pred == ICmpInst::ICMP_EQ && Br->getSuccessor(1) == LoopEntry))
304 return Cond->getOperand(0);
309 bool NclPopcountRecognize::detectIdiom(Instruction *&CntInst, PHINode *&CntPhi,
311 // Following code tries to detect this idiom:
314 // goto loop-exit // the precondition of the loop
317 // x1 = phi (x0, x2);
318 // cnt1 = phi(cnt0, cnt2);
322 // x2 = x1 & (x1 - 1);
329 // step 1: Check to see if the look-back branch match this pattern:
330 // "if (a!=0) goto loop-entry".
331 BasicBlock *LoopEntry;
332 Instruction *DefX2, *CountInst;
333 Value *VarX1, *VarX0;
334 PHINode *PhiX, *CountPhi;
336 DefX2 = CountInst = nullptr;
337 VarX1 = VarX0 = nullptr;
338 PhiX = CountPhi = nullptr;
339 LoopEntry = *(CurLoop->block_begin());
341 // step 1: Check if the loop-back branch is in desirable form.
343 if (Value *T = matchCondition(
344 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
345 DefX2 = dyn_cast<Instruction>(T);
350 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
352 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
355 BinaryOperator *SubOneOp;
357 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
358 VarX1 = DefX2->getOperand(1);
360 VarX1 = DefX2->getOperand(0);
361 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
366 Instruction *SubInst = cast<Instruction>(SubOneOp);
367 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
369 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
370 (SubInst->getOpcode() == Instruction::Add &&
371 Dec->isAllOnesValue()))) {
376 // step 3: Check the recurrence of variable X
378 PhiX = dyn_cast<PHINode>(VarX1);
380 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
385 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
388 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI(),
389 IterE = LoopEntry->end();
390 Iter != IterE; Iter++) {
391 Instruction *Inst = Iter;
392 if (Inst->getOpcode() != Instruction::Add)
395 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
396 if (!Inc || !Inc->isOne())
399 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
400 if (!Phi || Phi->getParent() != LoopEntry)
403 // Check if the result of the instruction is live of the loop.
404 bool LiveOutLoop = false;
405 for (User *U : Inst->users()) {
406 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
423 // step 5: check if the precondition is in this form:
424 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
426 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
427 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
428 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
439 void NclPopcountRecognize::transform(Instruction *CntInst, PHINode *CntPhi,
442 ScalarEvolution *SE = LIR.getScalarEvolution();
443 TargetLibraryInfo *TLI = LIR.getTargetLibraryInfo();
444 BasicBlock *PreHead = CurLoop->getLoopPreheader();
445 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
446 const DebugLoc DL = CntInst->getDebugLoc();
448 // Assuming before transformation, the loop is following:
449 // if (x) // the precondition
450 // do { cnt++; x &= x - 1; } while(x);
452 // Step 1: Insert the ctpop instruction at the end of the precondition block
453 IRBuilderTy Builder(PreCondBr);
454 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
456 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
457 NewCount = PopCntZext =
458 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
460 if (NewCount != PopCnt)
461 (cast<Instruction>(NewCount))->setDebugLoc(DL);
463 // TripCnt is exactly the number of iterations the loop has
466 // If the population counter's initial value is not zero, insert Add Inst.
467 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
468 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
469 if (!InitConst || !InitConst->isZero()) {
470 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
471 (cast<Instruction>(NewCount))->setDebugLoc(DL);
475 // Step 2: Replace the precondition from "if(x == 0) goto loop-exit" to
476 // "if(NewCount == 0) loop-exit". Withtout this change, the intrinsic
477 // function would be partial dead code, and downstream passes will drag
478 // it back from the precondition block to the preheader.
480 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
482 Value *Opnd0 = PopCntZext;
483 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
484 if (PreCond->getOperand(0) != Var)
485 std::swap(Opnd0, Opnd1);
487 ICmpInst *NewPreCond = cast<ICmpInst>(
488 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
489 PreCondBr->setCondition(NewPreCond);
491 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
494 // Step 3: Note that the population count is exactly the trip count of the
495 // loop in question, which enble us to to convert the loop from noncountable
496 // loop into a countable one. The benefit is twofold:
498 // - If the loop only counts population, the entire loop become dead after
499 // the transformation. It is lots easier to prove a countable loop dead
500 // than to prove a noncountable one. (In some C dialects, a infite loop
501 // isn't dead even if it computes nothing useful. In general, DCE needs
502 // to prove a noncountable loop finite before safely delete it.)
504 // - If the loop also performs something else, it remains alive.
505 // Since it is transformed to countable form, it can be aggressively
506 // optimized by some optimizations which are in general not applicable
507 // to a noncountable loop.
509 // After this step, this loop (conceptually) would look like following:
510 // newcnt = __builtin_ctpop(x);
513 // do { cnt++; x &= x-1; t--) } while (t > 0);
514 BasicBlock *Body = *(CurLoop->block_begin());
516 auto *LbBr = dyn_cast<BranchInst>(Body->getTerminator());
517 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
518 Type *Ty = TripCnt->getType();
520 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", Body->begin());
522 Builder.SetInsertPoint(LbCond);
523 Value *Opnd1 = cast<Value>(TcPhi);
524 Value *Opnd2 = cast<Value>(ConstantInt::get(Ty, 1));
525 Instruction *TcDec = cast<Instruction>(
526 Builder.CreateSub(Opnd1, Opnd2, "tcdec", false, true));
528 TcPhi->addIncoming(TripCnt, PreHead);
529 TcPhi->addIncoming(TcDec, Body);
531 CmpInst::Predicate Pred =
532 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
533 LbCond->setPredicate(Pred);
534 LbCond->setOperand(0, TcDec);
535 LbCond->setOperand(1, cast<Value>(ConstantInt::get(Ty, 0)));
538 // Step 4: All the references to the original population counter outside
539 // the loop are replaced with the NewCount -- the value returned from
540 // __builtin_ctpop().
541 CntInst->replaceUsesOutsideBlock(NewCount, Body);
543 // step 5: Forget the "non-computable" trip-count SCEV associated with the
544 // loop. The loop would otherwise not be deleted even if it becomes empty.
545 SE->forgetLoop(CurLoop);
548 CallInst *NclPopcountRecognize::createPopcntIntrinsic(IRBuilderTy &IRBuilder,
549 Value *Val, DebugLoc DL) {
550 Value *Ops[] = {Val};
551 Type *Tys[] = {Val->getType()};
553 Module *M = (*(CurLoop->block_begin()))->getParent()->getParent();
554 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
555 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
561 /// recognize - detect population count idiom in a non-countable loop. If
562 /// detected, transform the relevant code to popcount intrinsic function
563 /// call, and return true; otherwise, return false.
564 bool NclPopcountRecognize::recognize() {
565 if (!LIR.getTargetTransformInfo())
568 LIR.getScalarEvolution();
570 if (!preliminaryScreen())
573 Instruction *CntInst;
576 if (!detectIdiom(CntInst, CntPhi, Val))
579 transform(CntInst, CntPhi, Val);
583 //===----------------------------------------------------------------------===//
585 // Implementation of LoopIdiomRecognize
587 //===----------------------------------------------------------------------===//
589 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
590 if (skipOptnoneFunction(L))
595 // If the loop could not be converted to canonical form, it must have an
596 // indirectbr in it, just give up.
597 if (!L->getLoopPreheader())
600 // Disable loop idiom recognition if the function's name is a common idiom.
601 StringRef Name = L->getHeader()->getParent()->getName();
602 if (Name == "memset" || Name == "memcpy")
605 SE = &getAnalysis<ScalarEvolution>();
606 if (SE->hasLoopInvariantBackedgeTakenCount(L))
607 return runOnCountableLoop();
608 return runOnNoncountableLoop();
611 bool LoopIdiomRecognize::runOnCountableLoop() {
612 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
613 assert(!isa<SCEVCouldNotCompute>(BECount) &&
614 "runOnCountableLoop() called on a loop without a predictable"
615 "backedge-taken count");
617 // If this loop executes exactly one time, then it should be peeled, not
618 // optimized by this pass.
619 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
620 if (BECst->getValue()->getValue() == 0)
624 (void)getDominatorTree();
626 LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
627 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
630 (void)getTargetLibraryInfo();
632 SmallVector<BasicBlock *, 8> ExitBlocks;
633 CurLoop->getUniqueExitBlocks(ExitBlocks);
635 DEBUG(dbgs() << "loop-idiom Scanning: F["
636 << CurLoop->getHeader()->getParent()->getName() << "] Loop %"
637 << CurLoop->getHeader()->getName() << "\n");
639 bool MadeChange = false;
640 // Scan all the blocks in the loop that are not in subloops.
641 for (auto *BB : CurLoop->getBlocks()) {
642 // Ignore blocks in subloops.
643 if (LI.getLoopFor(BB) != CurLoop)
646 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
651 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
652 /// with the specified backedge count. This block is known to be in the current
653 /// loop and not in any subloops.
654 bool LoopIdiomRecognize::runOnLoopBlock(
655 BasicBlock *BB, const SCEV *BECount,
656 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
657 // We can only promote stores in this block if they are unconditionally
658 // executed in the loop. For a block to be unconditionally executed, it has
659 // to dominate all the exit blocks of the loop. Verify this now.
660 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
661 if (!DT->dominates(BB, ExitBlocks[i]))
664 bool MadeChange = false;
665 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
666 Instruction *Inst = I++;
667 // Look for store instructions, which may be optimized to memset/memcpy.
668 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
670 if (!processLoopStore(SI, BECount))
674 // If processing the store invalidated our iterator, start over from the
681 // Look for memset instructions, which may be optimized to a larger memset.
682 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
684 if (!processLoopMemSet(MSI, BECount))
688 // If processing the memset invalidated our iterator, start over from the
699 /// processLoopStore - See if this store can be promoted to a memset or memcpy.
700 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
704 Value *StoredVal = SI->getValueOperand();
705 Value *StorePtr = SI->getPointerOperand();
707 // Reject stores that are so large that they overflow an unsigned.
708 auto &DL = CurLoop->getHeader()->getModule()->getDataLayout();
709 uint64_t SizeInBits = DL.getTypeSizeInBits(StoredVal->getType());
710 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
713 // See if the pointer expression is an AddRec like {base,+,1} on the current
714 // loop, which indicates a strided store. If we have something else, it's a
715 // random store we can't handle.
716 const SCEVAddRecExpr *StoreEv =
717 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
718 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
721 // Check to see if the stride matches the size of the store. If so, then we
722 // know that every byte is touched in the loop.
723 unsigned StoreSize = (unsigned)SizeInBits >> 3;
724 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
726 if (!Stride || StoreSize != Stride->getValue()->getValue()) {
727 // TODO: Could also handle negative stride here someday, that will require
728 // the validity check in mayLoopAccessLocation to be updated though.
729 // Enable this to print exact negative strides.
730 if (0 && Stride && StoreSize == -Stride->getValue()->getValue()) {
731 dbgs() << "NEGATIVE STRIDE: " << *SI << "\n";
732 dbgs() << "BB: " << *SI->getParent();
738 // See if we can optimize just this store in isolation.
739 if (processLoopStridedStore(StorePtr, StoreSize, SI->getAlignment(),
740 StoredVal, SI, StoreEv, BECount))
743 // If the stored value is a strided load in the same loop with the same stride
744 // this this may be transformable into a memcpy. This kicks in for stuff like
745 // for (i) A[i] = B[i];
746 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
747 const SCEVAddRecExpr *LoadEv =
748 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
749 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
750 StoreEv->getOperand(1) == LoadEv->getOperand(1) && LI->isSimple())
751 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
754 // errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
759 /// processLoopMemSet - See if this memset can be promoted to a large memset.
760 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
761 const SCEV *BECount) {
762 // We can only handle non-volatile memsets with a constant size.
763 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength()))
766 // If we're not allowed to hack on memset, we fail.
767 if (!TLI->has(LibFunc::memset))
770 Value *Pointer = MSI->getDest();
772 // See if the pointer expression is an AddRec like {base,+,1} on the current
773 // loop, which indicates a strided store. If we have something else, it's a
774 // random store we can't handle.
775 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
776 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
779 // Reject memsets that are so large that they overflow an unsigned.
780 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
781 if ((SizeInBytes >> 32) != 0)
784 // Check to see if the stride matches the size of the memset. If so, then we
785 // know that every byte is touched in the loop.
786 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
788 // TODO: Could also handle negative stride here someday, that will require the
789 // validity check in mayLoopAccessLocation to be updated though.
790 if (!Stride || MSI->getLength() != Stride->getValue())
793 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
794 MSI->getAlignment(), MSI->getValue(), MSI, Ev,
798 /// mayLoopAccessLocation - Return true if the specified loop might access the
799 /// specified pointer location, which is a loop-strided access. The 'Access'
800 /// argument specifies what the verboten forms of access are (read or write).
801 static bool mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
802 const SCEV *BECount, unsigned StoreSize,
804 Instruction *IgnoredStore) {
805 // Get the location that may be stored across the loop. Since the access is
806 // strided positively through memory, we say that the modified location starts
807 // at the pointer and has infinite size.
808 uint64_t AccessSize = MemoryLocation::UnknownSize;
810 // If the loop iterates a fixed number of times, we can refine the access size
811 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
812 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
813 AccessSize = (BECst->getValue()->getZExtValue() + 1) * StoreSize;
815 // TODO: For this to be really effective, we have to dive into the pointer
816 // operand in the store. Store to &A[i] of 100 will always return may alias
817 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
818 // which will then no-alias a store to &A[100].
819 MemoryLocation StoreLoc(Ptr, AccessSize);
821 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
823 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
824 if (&*I != IgnoredStore && (AA.getModRefInfo(I, StoreLoc) & Access))
830 /// getMemSetPatternValue - If a strided store of the specified value is safe to
831 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
832 /// be passed in. Otherwise, return null.
834 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
835 /// just replicate their input array and then pass on to memset_pattern16.
836 static Constant *getMemSetPatternValue(Value *V, const DataLayout &DL) {
837 // If the value isn't a constant, we can't promote it to being in a constant
838 // array. We could theoretically do a store to an alloca or something, but
839 // that doesn't seem worthwhile.
840 Constant *C = dyn_cast<Constant>(V);
844 // Only handle simple values that are a power of two bytes in size.
845 uint64_t Size = DL.getTypeSizeInBits(V->getType());
846 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
849 // Don't care enough about darwin/ppc to implement this.
850 if (DL.isBigEndian())
853 // Convert to size in bytes.
856 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
857 // if the top and bottom are the same (e.g. for vectors and large integers).
861 // If the constant is exactly 16 bytes, just use it.
865 // Otherwise, we'll use an array of the constants.
866 unsigned ArraySize = 16 / Size;
867 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
868 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
871 /// processLoopStridedStore - We see a strided store of some value. If we can
872 /// transform this into a memset or memset_pattern in the loop preheader, do so.
873 bool LoopIdiomRecognize::processLoopStridedStore(
874 Value *DestPtr, unsigned StoreSize, unsigned StoreAlignment,
875 Value *StoredVal, Instruction *TheStore, const SCEVAddRecExpr *Ev,
876 const SCEV *BECount) {
878 // If the stored value is a byte-wise value (like i32 -1), then it may be
879 // turned into a memset of i8 -1, assuming that all the consecutive bytes
880 // are stored. A store of i32 0x01020304 can never be turned into a memset,
881 // but it can be turned into memset_pattern if the target supports it.
882 Value *SplatValue = isBytewiseValue(StoredVal);
883 Constant *PatternValue = nullptr;
884 auto &DL = CurLoop->getHeader()->getModule()->getDataLayout();
885 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
887 // If we're allowed to form a memset, and the stored value would be acceptable
888 // for memset, use it.
889 if (SplatValue && TLI->has(LibFunc::memset) &&
890 // Verify that the stored value is loop invariant. If not, we can't
891 // promote the memset.
892 CurLoop->isLoopInvariant(SplatValue)) {
893 // Keep and use SplatValue.
894 PatternValue = nullptr;
895 } else if (DestAS == 0 && TLI->has(LibFunc::memset_pattern16) &&
896 (PatternValue = getMemSetPatternValue(StoredVal, DL))) {
897 // Don't create memset_pattern16s with address spaces.
898 // It looks like we can use PatternValue!
899 SplatValue = nullptr;
901 // Otherwise, this isn't an idiom we can transform. For example, we can't
902 // do anything with a 3-byte store.
906 // The trip count of the loop and the base pointer of the addrec SCEV is
907 // guaranteed to be loop invariant, which means that it should dominate the
908 // header. This allows us to insert code for it in the preheader.
909 BasicBlock *Preheader = CurLoop->getLoopPreheader();
910 IRBuilder<> Builder(Preheader->getTerminator());
911 SCEVExpander Expander(*SE, DL, "loop-idiom");
913 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
915 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
916 // this into a memset in the loop preheader now if we want. However, this
917 // would be unsafe to do if there is anything else in the loop that may read
918 // or write to the aliased location. Check for any overlap by generating the
919 // base pointer and checking the region.
920 Value *BasePtr = Expander.expandCodeFor(Ev->getStart(), DestInt8PtrTy,
921 Preheader->getTerminator());
923 if (mayLoopAccessLocation(BasePtr, MRI_ModRef, CurLoop, BECount, StoreSize,
924 getAnalysis<AliasAnalysis>(), TheStore)) {
926 // If we generated new code for the base pointer, clean up.
927 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
931 // Okay, everything looks good, insert the memset.
933 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
934 // pointer size if it isn't already.
935 Type *IntPtr = Builder.getIntPtrTy(DL, DestAS);
936 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
938 const SCEV *NumBytesS =
939 SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1), SCEV::FlagNUW);
940 if (StoreSize != 1) {
941 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
946 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
951 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment);
953 // Everything is emitted in default address space
954 Type *Int8PtrTy = DestInt8PtrTy;
956 Module *M = TheStore->getParent()->getParent()->getParent();
958 M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(),
959 Int8PtrTy, Int8PtrTy, IntPtr, (void *)nullptr);
961 // Otherwise we should form a memset_pattern16. PatternValue is known to be
962 // an constant array of 16-bytes. Plop the value into a mergable global.
963 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
964 GlobalValue::PrivateLinkage,
965 PatternValue, ".memset_pattern");
966 GV->setUnnamedAddr(true); // Ok to merge these.
967 GV->setAlignment(16);
968 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
969 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
972 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
973 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
974 NewCall->setDebugLoc(TheStore->getDebugLoc());
976 // Okay, the memset has been formed. Zap the original store and anything that
978 deleteDeadInstruction(TheStore, TLI);
983 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
984 /// same-strided load.
985 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(
986 StoreInst *SI, unsigned StoreSize, const SCEVAddRecExpr *StoreEv,
987 const SCEVAddRecExpr *LoadEv, const SCEV *BECount) {
988 // If we're not allowed to form memcpy, we fail.
989 if (!TLI->has(LibFunc::memcpy))
992 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
994 // The trip count of the loop and the base pointer of the addrec SCEV is
995 // guaranteed to be loop invariant, which means that it should dominate the
996 // header. This allows us to insert code for it in the preheader.
997 BasicBlock *Preheader = CurLoop->getLoopPreheader();
998 IRBuilder<> Builder(Preheader->getTerminator());
999 const DataLayout &DL = Preheader->getModule()->getDataLayout();
1000 SCEVExpander Expander(*SE, DL, "loop-idiom");
1002 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
1003 // this into a memcpy in the loop preheader now if we want. However, this
1004 // would be unsafe to do if there is anything else in the loop that may read
1005 // or write the memory region we're storing to. This includes the load that
1006 // feeds the stores. Check for an alias by generating the base address and
1007 // checking everything.
1008 Value *StoreBasePtr = Expander.expandCodeFor(
1009 StoreEv->getStart(), Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
1010 Preheader->getTerminator());
1012 if (mayLoopAccessLocation(StoreBasePtr, MRI_ModRef, CurLoop, BECount,
1013 StoreSize, getAnalysis<AliasAnalysis>(), SI)) {
1015 // If we generated new code for the base pointer, clean up.
1016 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
1020 // For a memcpy, we have to make sure that the input array is not being
1021 // mutated by the loop.
1022 Value *LoadBasePtr = Expander.expandCodeFor(
1023 LoadEv->getStart(), Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
1024 Preheader->getTerminator());
1026 if (mayLoopAccessLocation(LoadBasePtr, MRI_Mod, CurLoop, BECount, StoreSize,
1027 getAnalysis<AliasAnalysis>(), SI)) {
1029 // If we generated new code for the base pointer, clean up.
1030 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
1031 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
1035 // Okay, everything is safe, we can transform this!
1037 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
1038 // pointer size if it isn't already.
1039 Type *IntPtrTy = Builder.getIntPtrTy(DL, SI->getPointerAddressSpace());
1040 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
1042 const SCEV *NumBytesS =
1043 SE->getAddExpr(BECount, SE->getConstant(IntPtrTy, 1), SCEV::FlagNUW);
1045 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
1049 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
1052 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
1053 std::min(SI->getAlignment(), LI->getAlignment()));
1054 NewCall->setDebugLoc(SI->getDebugLoc());
1056 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
1057 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
1058 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
1060 // Okay, the memset has been formed. Zap the original store and anything that
1062 deleteDeadInstruction(SI, TLI);
1067 bool LoopIdiomRecognize::runOnNoncountableLoop() {
1068 NclPopcountRecognize Popcount(*this);
1069 if (Popcount.recognize())