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 // If compiling for code size we avoid idiom recognition if the resulting
15 // code could be larger than the code for the original loop. One way this could
16 // happen is if the loop is not removable after idiom recognition due to the
17 // presence of non-idiom instructions. The initial implementation of the
18 // heuristics applies to idioms in multi-block loops.
20 //===----------------------------------------------------------------------===//
24 // Future loop memory idioms to recognize:
25 // memcmp, memmove, strlen, etc.
26 // Future floating point idioms to recognize in -ffast-math mode:
28 // Future integer operation idioms to recognize:
31 // Beware that isel's default lowering for ctpop is highly inefficient for
32 // i64 and larger types when i64 is legal and the value has few bits set. It
33 // would be good to enhance isel to emit a loop for ctpop in this case.
35 // This could recognize common matrix multiplies and dot product idioms and
36 // replace them with calls to BLAS (if linked in??).
38 //===----------------------------------------------------------------------===//
40 #include "llvm/Transforms/Scalar/LoopIdiomRecognize.h"
41 #include "llvm/ADT/MapVector.h"
42 #include "llvm/ADT/SetVector.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/Analysis/AliasAnalysis.h"
45 #include "llvm/Analysis/BasicAliasAnalysis.h"
46 #include "llvm/Analysis/GlobalsModRef.h"
47 #include "llvm/Analysis/LoopAccessAnalysis.h"
48 #include "llvm/Analysis/LoopPass.h"
49 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
50 #include "llvm/Analysis/ScalarEvolutionExpander.h"
51 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
52 #include "llvm/Analysis/TargetLibraryInfo.h"
53 #include "llvm/Analysis/TargetTransformInfo.h"
54 #include "llvm/Analysis/ValueTracking.h"
55 #include "llvm/IR/DataLayout.h"
56 #include "llvm/IR/Dominators.h"
57 #include "llvm/IR/IRBuilder.h"
58 #include "llvm/IR/IntrinsicInst.h"
59 #include "llvm/IR/Module.h"
60 #include "llvm/Support/Debug.h"
61 #include "llvm/Support/raw_ostream.h"
62 #include "llvm/Transforms/Scalar.h"
63 #include "llvm/Transforms/Scalar/LoopPassManager.h"
64 #include "llvm/Transforms/Utils/BuildLibCalls.h"
65 #include "llvm/Transforms/Utils/Local.h"
66 #include "llvm/Transforms/Utils/LoopUtils.h"
69 #define DEBUG_TYPE "loop-idiom"
71 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
72 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
74 static cl::opt<bool> UseLIRCodeSizeHeurs(
75 "use-lir-code-size-heurs",
76 cl::desc("Use loop idiom recognition code size heuristics when compiling"
78 cl::init(true), cl::Hidden);
82 class LoopIdiomRecognize {
88 TargetLibraryInfo *TLI;
89 const TargetTransformInfo *TTI;
91 bool ApplyCodeSizeHeuristics;
94 explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT,
95 LoopInfo *LI, ScalarEvolution *SE,
96 TargetLibraryInfo *TLI,
97 const TargetTransformInfo *TTI,
99 : CurLoop(nullptr), AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI),
102 bool runOnLoop(Loop *L);
105 typedef SmallVector<StoreInst *, 8> StoreList;
106 typedef MapVector<Value *, StoreList> StoreListMap;
107 StoreListMap StoreRefsForMemset;
108 StoreListMap StoreRefsForMemsetPattern;
109 StoreList StoreRefsForMemcpy;
111 bool HasMemsetPattern;
113 /// Return code for isLegalStore()
114 enum LegalStoreKind {
119 DontUse // Dummy retval never to be used. Allows catching errors in retval
123 /// \name Countable Loop Idiom Handling
126 bool runOnCountableLoop();
127 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
128 SmallVectorImpl<BasicBlock *> &ExitBlocks);
130 void collectStores(BasicBlock *BB);
131 LegalStoreKind isLegalStore(StoreInst *SI);
132 bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount,
134 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
136 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
137 unsigned StoreAlignment, Value *StoredVal,
138 Instruction *TheStore,
139 SmallPtrSetImpl<Instruction *> &Stores,
140 const SCEVAddRecExpr *Ev, const SCEV *BECount,
141 bool NegStride, bool IsLoopMemset = false);
142 bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount);
143 bool avoidLIRForMultiBlockLoop(bool IsMemset = false,
144 bool IsLoopMemset = false);
147 /// \name Noncountable Loop Idiom Handling
150 bool runOnNoncountableLoop();
152 bool recognizePopcount();
153 void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
154 PHINode *CntPhi, Value *Var);
155 bool recognizeAndInsertCTLZ();
156 void transformLoopToCountable(BasicBlock *PreCondBB, Instruction *CntInst,
157 PHINode *CntPhi, Value *Var, const DebugLoc DL,
158 bool ZeroCheck, bool IsCntPhiUsedOutsideLoop);
163 class LoopIdiomRecognizeLegacyPass : public LoopPass {
166 explicit LoopIdiomRecognizeLegacyPass() : LoopPass(ID) {
167 initializeLoopIdiomRecognizeLegacyPassPass(
168 *PassRegistry::getPassRegistry());
171 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
175 AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
176 DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
177 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
178 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
179 TargetLibraryInfo *TLI =
180 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
181 const TargetTransformInfo *TTI =
182 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
183 *L->getHeader()->getParent());
184 const DataLayout *DL = &L->getHeader()->getModule()->getDataLayout();
186 LoopIdiomRecognize LIR(AA, DT, LI, SE, TLI, TTI, DL);
187 return LIR.runOnLoop(L);
190 /// This transformation requires natural loop information & requires that
191 /// loop preheaders be inserted into the CFG.
193 void getAnalysisUsage(AnalysisUsage &AU) const override {
194 AU.addRequired<TargetLibraryInfoWrapperPass>();
195 AU.addRequired<TargetTransformInfoWrapperPass>();
196 getLoopAnalysisUsage(AU);
199 } // End anonymous namespace.
201 PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM,
202 LoopStandardAnalysisResults &AR,
204 const auto *DL = &L.getHeader()->getModule()->getDataLayout();
206 LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI, DL);
207 if (!LIR.runOnLoop(&L))
208 return PreservedAnalyses::all();
210 return getLoopPassPreservedAnalyses();
213 char LoopIdiomRecognizeLegacyPass::ID = 0;
214 INITIALIZE_PASS_BEGIN(LoopIdiomRecognizeLegacyPass, "loop-idiom",
215 "Recognize loop idioms", false, false)
216 INITIALIZE_PASS_DEPENDENCY(LoopPass)
217 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
218 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
219 INITIALIZE_PASS_END(LoopIdiomRecognizeLegacyPass, "loop-idiom",
220 "Recognize loop idioms", false, false)
222 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognizeLegacyPass(); }
224 static void deleteDeadInstruction(Instruction *I) {
225 I->replaceAllUsesWith(UndefValue::get(I->getType()));
226 I->eraseFromParent();
229 //===----------------------------------------------------------------------===//
231 // Implementation of LoopIdiomRecognize
233 //===----------------------------------------------------------------------===//
235 bool LoopIdiomRecognize::runOnLoop(Loop *L) {
237 // If the loop could not be converted to canonical form, it must have an
238 // indirectbr in it, just give up.
239 if (!L->getLoopPreheader())
242 // Disable loop idiom recognition if the function's name is a common idiom.
243 StringRef Name = L->getHeader()->getParent()->getName();
244 if (Name == "memset" || Name == "memcpy")
247 // Determine if code size heuristics need to be applied.
248 ApplyCodeSizeHeuristics =
249 L->getHeader()->getParent()->optForSize() && UseLIRCodeSizeHeurs;
251 HasMemset = TLI->has(LibFunc_memset);
252 HasMemsetPattern = TLI->has(LibFunc_memset_pattern16);
253 HasMemcpy = TLI->has(LibFunc_memcpy);
255 if (HasMemset || HasMemsetPattern || HasMemcpy)
256 if (SE->hasLoopInvariantBackedgeTakenCount(L))
257 return runOnCountableLoop();
259 return runOnNoncountableLoop();
262 bool LoopIdiomRecognize::runOnCountableLoop() {
263 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
264 assert(!isa<SCEVCouldNotCompute>(BECount) &&
265 "runOnCountableLoop() called on a loop without a predictable"
266 "backedge-taken count");
268 // If this loop executes exactly one time, then it should be peeled, not
269 // optimized by this pass.
270 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
271 if (BECst->getAPInt() == 0)
274 SmallVector<BasicBlock *, 8> ExitBlocks;
275 CurLoop->getUniqueExitBlocks(ExitBlocks);
277 DEBUG(dbgs() << "loop-idiom Scanning: F["
278 << CurLoop->getHeader()->getParent()->getName() << "] Loop %"
279 << CurLoop->getHeader()->getName() << "\n");
281 bool MadeChange = false;
283 // The following transforms hoist stores/memsets into the loop pre-header.
284 // Give up if the loop has instructions may throw.
285 LoopSafetyInfo SafetyInfo;
286 computeLoopSafetyInfo(&SafetyInfo, CurLoop);
287 if (SafetyInfo.MayThrow)
290 // Scan all the blocks in the loop that are not in subloops.
291 for (auto *BB : CurLoop->getBlocks()) {
292 // Ignore blocks in subloops.
293 if (LI->getLoopFor(BB) != CurLoop)
296 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
301 static unsigned getStoreSizeInBytes(StoreInst *SI, const DataLayout *DL) {
302 uint64_t SizeInBits = DL->getTypeSizeInBits(SI->getValueOperand()->getType());
303 assert(((SizeInBits & 7) || (SizeInBits >> 32) == 0) &&
304 "Don't overflow unsigned.");
305 return (unsigned)SizeInBits >> 3;
308 static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) {
309 const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1));
310 return ConstStride->getAPInt();
313 /// getMemSetPatternValue - If a strided store of the specified value is safe to
314 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
315 /// be passed in. Otherwise, return null.
317 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
318 /// just replicate their input array and then pass on to memset_pattern16.
319 static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) {
320 // If the value isn't a constant, we can't promote it to being in a constant
321 // array. We could theoretically do a store to an alloca or something, but
322 // that doesn't seem worthwhile.
323 Constant *C = dyn_cast<Constant>(V);
327 // Only handle simple values that are a power of two bytes in size.
328 uint64_t Size = DL->getTypeSizeInBits(V->getType());
329 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
332 // Don't care enough about darwin/ppc to implement this.
333 if (DL->isBigEndian())
336 // Convert to size in bytes.
339 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
340 // if the top and bottom are the same (e.g. for vectors and large integers).
344 // If the constant is exactly 16 bytes, just use it.
348 // Otherwise, we'll use an array of the constants.
349 unsigned ArraySize = 16 / Size;
350 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
351 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
354 LoopIdiomRecognize::LegalStoreKind
355 LoopIdiomRecognize::isLegalStore(StoreInst *SI) {
356 // Don't touch volatile stores.
358 return LegalStoreKind::None;
360 // Don't convert stores of non-integral pointer types to memsets (which stores
362 if (DL->isNonIntegralPointerType(SI->getValueOperand()->getType()))
363 return LegalStoreKind::None;
365 // Avoid merging nontemporal stores.
366 if (SI->getMetadata(LLVMContext::MD_nontemporal))
367 return LegalStoreKind::None;
369 Value *StoredVal = SI->getValueOperand();
370 Value *StorePtr = SI->getPointerOperand();
372 // Reject stores that are so large that they overflow an unsigned.
373 uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
374 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
375 return LegalStoreKind::None;
377 // See if the pointer expression is an AddRec like {base,+,1} on the current
378 // loop, which indicates a strided store. If we have something else, it's a
379 // random store we can't handle.
380 const SCEVAddRecExpr *StoreEv =
381 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
382 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
383 return LegalStoreKind::None;
385 // Check to see if we have a constant stride.
386 if (!isa<SCEVConstant>(StoreEv->getOperand(1)))
387 return LegalStoreKind::None;
389 // See if the store can be turned into a memset.
391 // If the stored value is a byte-wise value (like i32 -1), then it may be
392 // turned into a memset of i8 -1, assuming that all the consecutive bytes
393 // are stored. A store of i32 0x01020304 can never be turned into a memset,
394 // but it can be turned into memset_pattern if the target supports it.
395 Value *SplatValue = isBytewiseValue(StoredVal);
396 Constant *PatternValue = nullptr;
398 // If we're allowed to form a memset, and the stored value would be
399 // acceptable for memset, use it.
400 if (HasMemset && SplatValue &&
401 // Verify that the stored value is loop invariant. If not, we can't
402 // promote the memset.
403 CurLoop->isLoopInvariant(SplatValue)) {
404 // It looks like we can use SplatValue.
405 return LegalStoreKind::Memset;
406 } else if (HasMemsetPattern &&
407 // Don't create memset_pattern16s with address spaces.
408 StorePtr->getType()->getPointerAddressSpace() == 0 &&
409 (PatternValue = getMemSetPatternValue(StoredVal, DL))) {
410 // It looks like we can use PatternValue!
411 return LegalStoreKind::MemsetPattern;
414 // Otherwise, see if the store can be turned into a memcpy.
416 // Check to see if the stride matches the size of the store. If so, then we
417 // know that every byte is touched in the loop.
418 APInt Stride = getStoreStride(StoreEv);
419 unsigned StoreSize = getStoreSizeInBytes(SI, DL);
420 if (StoreSize != Stride && StoreSize != -Stride)
421 return LegalStoreKind::None;
423 // The store must be feeding a non-volatile load.
424 LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand());
425 if (!LI || !LI->isSimple())
426 return LegalStoreKind::None;
428 // See if the pointer expression is an AddRec like {base,+,1} on the current
429 // loop, which indicates a strided load. If we have something else, it's a
430 // random load we can't handle.
431 const SCEVAddRecExpr *LoadEv =
432 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
433 if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
434 return LegalStoreKind::None;
436 // The store and load must share the same stride.
437 if (StoreEv->getOperand(1) != LoadEv->getOperand(1))
438 return LegalStoreKind::None;
440 // Success. This store can be converted into a memcpy.
441 return LegalStoreKind::Memcpy;
443 // This store can't be transformed into a memset/memcpy.
444 return LegalStoreKind::None;
447 void LoopIdiomRecognize::collectStores(BasicBlock *BB) {
448 StoreRefsForMemset.clear();
449 StoreRefsForMemsetPattern.clear();
450 StoreRefsForMemcpy.clear();
451 for (Instruction &I : *BB) {
452 StoreInst *SI = dyn_cast<StoreInst>(&I);
456 // Make sure this is a strided store with a constant stride.
457 switch (isLegalStore(SI)) {
458 case LegalStoreKind::None:
461 case LegalStoreKind::Memset: {
462 // Find the base pointer.
463 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), *DL);
464 StoreRefsForMemset[Ptr].push_back(SI);
466 case LegalStoreKind::MemsetPattern: {
467 // Find the base pointer.
468 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), *DL);
469 StoreRefsForMemsetPattern[Ptr].push_back(SI);
471 case LegalStoreKind::Memcpy:
472 StoreRefsForMemcpy.push_back(SI);
475 assert(false && "unhandled return value");
481 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
482 /// with the specified backedge count. This block is known to be in the current
483 /// loop and not in any subloops.
484 bool LoopIdiomRecognize::runOnLoopBlock(
485 BasicBlock *BB, const SCEV *BECount,
486 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
487 // We can only promote stores in this block if they are unconditionally
488 // executed in the loop. For a block to be unconditionally executed, it has
489 // to dominate all the exit blocks of the loop. Verify this now.
490 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
491 if (!DT->dominates(BB, ExitBlocks[i]))
494 bool MadeChange = false;
495 // Look for store instructions, which may be optimized to memset/memcpy.
498 // Look for a single store or sets of stores with a common base, which can be
499 // optimized into a memset (memset_pattern). The latter most commonly happens
500 // with structs and handunrolled loops.
501 for (auto &SL : StoreRefsForMemset)
502 MadeChange |= processLoopStores(SL.second, BECount, true);
504 for (auto &SL : StoreRefsForMemsetPattern)
505 MadeChange |= processLoopStores(SL.second, BECount, false);
507 // Optimize the store into a memcpy, if it feeds an similarly strided load.
508 for (auto &SI : StoreRefsForMemcpy)
509 MadeChange |= processLoopStoreOfLoopLoad(SI, BECount);
511 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
512 Instruction *Inst = &*I++;
513 // Look for memset instructions, which may be optimized to a larger memset.
514 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
515 WeakTrackingVH InstPtr(&*I);
516 if (!processLoopMemSet(MSI, BECount))
520 // If processing the memset invalidated our iterator, start over from the
531 /// processLoopStores - See if this store(s) can be promoted to a memset.
532 bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL,
535 // Try to find consecutive stores that can be transformed into memsets.
536 SetVector<StoreInst *> Heads, Tails;
537 SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain;
539 // Do a quadratic search on all of the given stores and find
540 // all of the pairs of stores that follow each other.
541 SmallVector<unsigned, 16> IndexQueue;
542 for (unsigned i = 0, e = SL.size(); i < e; ++i) {
543 assert(SL[i]->isSimple() && "Expected only non-volatile stores.");
545 Value *FirstStoredVal = SL[i]->getValueOperand();
546 Value *FirstStorePtr = SL[i]->getPointerOperand();
547 const SCEVAddRecExpr *FirstStoreEv =
548 cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr));
549 APInt FirstStride = getStoreStride(FirstStoreEv);
550 unsigned FirstStoreSize = getStoreSizeInBytes(SL[i], DL);
552 // See if we can optimize just this store in isolation.
553 if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) {
558 Value *FirstSplatValue = nullptr;
559 Constant *FirstPatternValue = nullptr;
562 FirstSplatValue = isBytewiseValue(FirstStoredVal);
564 FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL);
566 assert((FirstSplatValue || FirstPatternValue) &&
567 "Expected either splat value or pattern value.");
570 // If a store has multiple consecutive store candidates, search Stores
571 // array according to the sequence: from i+1 to e, then from i-1 to 0.
572 // This is because usually pairing with immediate succeeding or preceding
573 // candidate create the best chance to find memset opportunity.
575 for (j = i + 1; j < e; ++j)
576 IndexQueue.push_back(j);
577 for (j = i; j > 0; --j)
578 IndexQueue.push_back(j - 1);
580 for (auto &k : IndexQueue) {
581 assert(SL[k]->isSimple() && "Expected only non-volatile stores.");
582 Value *SecondStorePtr = SL[k]->getPointerOperand();
583 const SCEVAddRecExpr *SecondStoreEv =
584 cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr));
585 APInt SecondStride = getStoreStride(SecondStoreEv);
587 if (FirstStride != SecondStride)
590 Value *SecondStoredVal = SL[k]->getValueOperand();
591 Value *SecondSplatValue = nullptr;
592 Constant *SecondPatternValue = nullptr;
595 SecondSplatValue = isBytewiseValue(SecondStoredVal);
597 SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL);
599 assert((SecondSplatValue || SecondPatternValue) &&
600 "Expected either splat value or pattern value.");
602 if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) {
604 if (FirstSplatValue != SecondSplatValue)
607 if (FirstPatternValue != SecondPatternValue)
612 ConsecutiveChain[SL[i]] = SL[k];
618 // We may run into multiple chains that merge into a single chain. We mark the
619 // stores that we transformed so that we don't visit the same store twice.
620 SmallPtrSet<Value *, 16> TransformedStores;
621 bool Changed = false;
623 // For stores that start but don't end a link in the chain:
624 for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end();
626 if (Tails.count(*it))
629 // We found a store instr that starts a chain. Now follow the chain and try
631 SmallPtrSet<Instruction *, 8> AdjacentStores;
634 StoreInst *HeadStore = I;
635 unsigned StoreSize = 0;
637 // Collect the chain into a list.
638 while (Tails.count(I) || Heads.count(I)) {
639 if (TransformedStores.count(I))
641 AdjacentStores.insert(I);
643 StoreSize += getStoreSizeInBytes(I, DL);
644 // Move to the next value in the chain.
645 I = ConsecutiveChain[I];
648 Value *StoredVal = HeadStore->getValueOperand();
649 Value *StorePtr = HeadStore->getPointerOperand();
650 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
651 APInt Stride = getStoreStride(StoreEv);
653 // Check to see if the stride matches the size of the stores. If so, then
654 // we know that every byte is touched in the loop.
655 if (StoreSize != Stride && StoreSize != -Stride)
658 bool NegStride = StoreSize == -Stride;
660 if (processLoopStridedStore(StorePtr, StoreSize, HeadStore->getAlignment(),
661 StoredVal, HeadStore, AdjacentStores, StoreEv,
662 BECount, NegStride)) {
663 TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end());
671 /// processLoopMemSet - See if this memset can be promoted to a large memset.
672 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
673 const SCEV *BECount) {
674 // We can only handle non-volatile memsets with a constant size.
675 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength()))
678 // If we're not allowed to hack on memset, we fail.
682 Value *Pointer = MSI->getDest();
684 // See if the pointer expression is an AddRec like {base,+,1} on the current
685 // loop, which indicates a strided store. If we have something else, it's a
686 // random store we can't handle.
687 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
688 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
691 // Reject memsets that are so large that they overflow an unsigned.
692 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
693 if ((SizeInBytes >> 32) != 0)
696 // Check to see if the stride matches the size of the memset. If so, then we
697 // know that every byte is touched in the loop.
698 const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
702 APInt Stride = ConstStride->getAPInt();
703 if (SizeInBytes != Stride && SizeInBytes != -Stride)
706 // Verify that the memset value is loop invariant. If not, we can't promote
708 Value *SplatValue = MSI->getValue();
709 if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue))
712 SmallPtrSet<Instruction *, 1> MSIs;
714 bool NegStride = SizeInBytes == -Stride;
715 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
716 MSI->getAlignment(), SplatValue, MSI, MSIs, Ev,
717 BECount, NegStride, /*IsLoopMemset=*/true);
720 /// mayLoopAccessLocation - Return true if the specified loop might access the
721 /// specified pointer location, which is a loop-strided access. The 'Access'
722 /// argument specifies what the verboten forms of access are (read or write).
724 mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
725 const SCEV *BECount, unsigned StoreSize,
727 SmallPtrSetImpl<Instruction *> &IgnoredStores) {
728 // Get the location that may be stored across the loop. Since the access is
729 // strided positively through memory, we say that the modified location starts
730 // at the pointer and has infinite size.
731 uint64_t AccessSize = MemoryLocation::UnknownSize;
733 // If the loop iterates a fixed number of times, we can refine the access size
734 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
735 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
736 AccessSize = (BECst->getValue()->getZExtValue() + 1) * StoreSize;
738 // TODO: For this to be really effective, we have to dive into the pointer
739 // operand in the store. Store to &A[i] of 100 will always return may alias
740 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
741 // which will then no-alias a store to &A[100].
742 MemoryLocation StoreLoc(Ptr, AccessSize);
744 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
746 for (Instruction &I : **BI)
747 if (IgnoredStores.count(&I) == 0 &&
748 (AA.getModRefInfo(&I, StoreLoc) & Access))
754 // If we have a negative stride, Start refers to the end of the memory location
755 // we're trying to memset. Therefore, we need to recompute the base pointer,
756 // which is just Start - BECount*Size.
757 static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount,
758 Type *IntPtr, unsigned StoreSize,
759 ScalarEvolution *SE) {
760 const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr);
762 Index = SE->getMulExpr(Index, SE->getConstant(IntPtr, StoreSize),
764 return SE->getMinusSCEV(Start, Index);
767 /// processLoopStridedStore - We see a strided store of some value. If we can
768 /// transform this into a memset or memset_pattern in the loop preheader, do so.
769 bool LoopIdiomRecognize::processLoopStridedStore(
770 Value *DestPtr, unsigned StoreSize, unsigned StoreAlignment,
771 Value *StoredVal, Instruction *TheStore,
772 SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev,
773 const SCEV *BECount, bool NegStride, bool IsLoopMemset) {
774 Value *SplatValue = isBytewiseValue(StoredVal);
775 Constant *PatternValue = nullptr;
778 PatternValue = getMemSetPatternValue(StoredVal, DL);
780 assert((SplatValue || PatternValue) &&
781 "Expected either splat value or pattern value.");
783 // The trip count of the loop and the base pointer of the addrec SCEV is
784 // guaranteed to be loop invariant, which means that it should dominate the
785 // header. This allows us to insert code for it in the preheader.
786 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
787 BasicBlock *Preheader = CurLoop->getLoopPreheader();
788 IRBuilder<> Builder(Preheader->getTerminator());
789 SCEVExpander Expander(*SE, *DL, "loop-idiom");
791 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
792 Type *IntPtr = Builder.getIntPtrTy(*DL, DestAS);
794 const SCEV *Start = Ev->getStart();
795 // Handle negative strided loops.
797 Start = getStartForNegStride(Start, BECount, IntPtr, StoreSize, SE);
799 // TODO: ideally we should still be able to generate memset if SCEV expander
800 // is taught to generate the dependencies at the latest point.
801 if (!isSafeToExpand(Start, *SE))
804 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
805 // this into a memset in the loop preheader now if we want. However, this
806 // would be unsafe to do if there is anything else in the loop that may read
807 // or write to the aliased location. Check for any overlap by generating the
808 // base pointer and checking the region.
810 Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator());
811 if (mayLoopAccessLocation(BasePtr, MRI_ModRef, CurLoop, BECount, StoreSize,
814 // If we generated new code for the base pointer, clean up.
815 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
819 if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset))
822 // Okay, everything looks good, insert the memset.
824 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
825 // pointer size if it isn't already.
826 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
828 const SCEV *NumBytesS =
829 SE->getAddExpr(BECount, SE->getOne(IntPtr), SCEV::FlagNUW);
830 if (StoreSize != 1) {
831 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
835 // TODO: ideally we should still be able to generate memset if SCEV expander
836 // is taught to generate the dependencies at the latest point.
837 if (!isSafeToExpand(NumBytesS, *SE))
841 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
846 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment);
848 // Everything is emitted in default address space
849 Type *Int8PtrTy = DestInt8PtrTy;
851 Module *M = TheStore->getModule();
853 M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(),
854 Int8PtrTy, Int8PtrTy, IntPtr);
855 inferLibFuncAttributes(*M->getFunction("memset_pattern16"), *TLI);
857 // Otherwise we should form a memset_pattern16. PatternValue is known to be
858 // an constant array of 16-bytes. Plop the value into a mergable global.
859 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
860 GlobalValue::PrivateLinkage,
861 PatternValue, ".memset_pattern");
862 GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these.
863 GV->setAlignment(16);
864 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
865 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
868 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
869 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
870 NewCall->setDebugLoc(TheStore->getDebugLoc());
872 // Okay, the memset has been formed. Zap the original store and anything that
874 for (auto *I : Stores)
875 deleteDeadInstruction(I);
880 /// If the stored value is a strided load in the same loop with the same stride
881 /// this may be transformable into a memcpy. This kicks in for stuff like
882 /// for (i) A[i] = B[i];
883 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI,
884 const SCEV *BECount) {
885 assert(SI->isSimple() && "Expected only non-volatile stores.");
887 Value *StorePtr = SI->getPointerOperand();
888 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
889 APInt Stride = getStoreStride(StoreEv);
890 unsigned StoreSize = getStoreSizeInBytes(SI, DL);
891 bool NegStride = StoreSize == -Stride;
893 // The store must be feeding a non-volatile load.
894 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
895 assert(LI->isSimple() && "Expected only non-volatile stores.");
897 // See if the pointer expression is an AddRec like {base,+,1} on the current
898 // loop, which indicates a strided load. If we have something else, it's a
899 // random load we can't handle.
900 const SCEVAddRecExpr *LoadEv =
901 cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
903 // The trip count of the loop and the base pointer of the addrec SCEV is
904 // guaranteed to be loop invariant, which means that it should dominate the
905 // header. This allows us to insert code for it in the preheader.
906 BasicBlock *Preheader = CurLoop->getLoopPreheader();
907 IRBuilder<> Builder(Preheader->getTerminator());
908 SCEVExpander Expander(*SE, *DL, "loop-idiom");
910 const SCEV *StrStart = StoreEv->getStart();
911 unsigned StrAS = SI->getPointerAddressSpace();
912 Type *IntPtrTy = Builder.getIntPtrTy(*DL, StrAS);
914 // Handle negative strided loops.
916 StrStart = getStartForNegStride(StrStart, BECount, IntPtrTy, StoreSize, SE);
918 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
919 // this into a memcpy in the loop preheader now if we want. However, this
920 // would be unsafe to do if there is anything else in the loop that may read
921 // or write the memory region we're storing to. This includes the load that
922 // feeds the stores. Check for an alias by generating the base address and
923 // checking everything.
924 Value *StoreBasePtr = Expander.expandCodeFor(
925 StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator());
927 SmallPtrSet<Instruction *, 1> Stores;
929 if (mayLoopAccessLocation(StoreBasePtr, MRI_ModRef, CurLoop, BECount,
930 StoreSize, *AA, Stores)) {
932 // If we generated new code for the base pointer, clean up.
933 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
937 const SCEV *LdStart = LoadEv->getStart();
938 unsigned LdAS = LI->getPointerAddressSpace();
940 // Handle negative strided loops.
942 LdStart = getStartForNegStride(LdStart, BECount, IntPtrTy, StoreSize, SE);
944 // For a memcpy, we have to make sure that the input array is not being
945 // mutated by the loop.
946 Value *LoadBasePtr = Expander.expandCodeFor(
947 LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator());
949 if (mayLoopAccessLocation(LoadBasePtr, MRI_Mod, CurLoop, BECount, StoreSize,
952 // If we generated new code for the base pointer, clean up.
953 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
954 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
958 if (avoidLIRForMultiBlockLoop())
961 // Okay, everything is safe, we can transform this!
963 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
964 // pointer size if it isn't already.
965 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
967 const SCEV *NumBytesS =
968 SE->getAddExpr(BECount, SE->getOne(IntPtrTy), SCEV::FlagNUW);
970 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
974 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
977 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
978 std::min(SI->getAlignment(), LI->getAlignment()));
979 NewCall->setDebugLoc(SI->getDebugLoc());
981 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
982 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
983 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
985 // Okay, the memcpy has been formed. Zap the original store and anything that
987 deleteDeadInstruction(SI);
992 // When compiling for codesize we avoid idiom recognition for a multi-block loop
993 // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop.
995 bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset,
997 if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) {
998 if (!CurLoop->getParentLoop() && (!IsMemset || !IsLoopMemset)) {
999 DEBUG(dbgs() << " " << CurLoop->getHeader()->getParent()->getName()
1000 << " : LIR " << (IsMemset ? "Memset" : "Memcpy")
1001 << " avoided: multi-block top-level loop\n");
1009 bool LoopIdiomRecognize::runOnNoncountableLoop() {
1010 return recognizePopcount() || recognizeAndInsertCTLZ();
1013 /// Check if the given conditional branch is based on the comparison between
1014 /// a variable and zero, and if the variable is non-zero, the control yields to
1015 /// the loop entry. If the branch matches the behavior, the variable involved
1016 /// in the comparison is returned. This function will be called to see if the
1017 /// precondition and postcondition of the loop are in desirable form.
1018 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry) {
1019 if (!BI || !BI->isConditional())
1022 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1026 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
1027 if (!CmpZero || !CmpZero->isZero())
1030 ICmpInst::Predicate Pred = Cond->getPredicate();
1031 if ((Pred == ICmpInst::ICMP_NE && BI->getSuccessor(0) == LoopEntry) ||
1032 (Pred == ICmpInst::ICMP_EQ && BI->getSuccessor(1) == LoopEntry))
1033 return Cond->getOperand(0);
1038 // Check if the recurrence variable `VarX` is in the right form to create
1039 // the idiom. Returns the value coerced to a PHINode if so.
1040 static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX,
1041 BasicBlock *LoopEntry) {
1042 auto *PhiX = dyn_cast<PHINode>(VarX);
1043 if (PhiX && PhiX->getParent() == LoopEntry &&
1044 (PhiX->getOperand(0) == DefX || PhiX->getOperand(1) == DefX))
1049 /// Return true iff the idiom is detected in the loop.
1052 /// 1) \p CntInst is set to the instruction counting the population bit.
1053 /// 2) \p CntPhi is set to the corresponding phi node.
1054 /// 3) \p Var is set to the value whose population bits are being counted.
1056 /// The core idiom we are trying to detect is:
1059 /// goto loop-exit // the precondition of the loop
1060 /// cnt0 = init-val;
1062 /// x1 = phi (x0, x2);
1063 /// cnt1 = phi(cnt0, cnt2);
1065 /// cnt2 = cnt1 + 1;
1067 /// x2 = x1 & (x1 - 1);
1069 /// } while(x != 0);
1073 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
1074 Instruction *&CntInst, PHINode *&CntPhi,
1076 // step 1: Check to see if the look-back branch match this pattern:
1077 // "if (a!=0) goto loop-entry".
1078 BasicBlock *LoopEntry;
1079 Instruction *DefX2, *CountInst;
1080 Value *VarX1, *VarX0;
1081 PHINode *PhiX, *CountPhi;
1083 DefX2 = CountInst = nullptr;
1084 VarX1 = VarX0 = nullptr;
1085 PhiX = CountPhi = nullptr;
1086 LoopEntry = *(CurLoop->block_begin());
1088 // step 1: Check if the loop-back branch is in desirable form.
1090 if (Value *T = matchCondition(
1091 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1092 DefX2 = dyn_cast<Instruction>(T);
1097 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
1099 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
1102 BinaryOperator *SubOneOp;
1104 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
1105 VarX1 = DefX2->getOperand(1);
1107 VarX1 = DefX2->getOperand(0);
1108 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
1113 Instruction *SubInst = cast<Instruction>(SubOneOp);
1114 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
1116 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
1117 (SubInst->getOpcode() == Instruction::Add &&
1118 Dec->isAllOnesValue()))) {
1123 // step 3: Check the recurrence of variable X
1124 PhiX = getRecurrenceVar(VarX1, DefX2, LoopEntry);
1128 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
1130 CountInst = nullptr;
1131 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
1132 IterE = LoopEntry->end();
1133 Iter != IterE; Iter++) {
1134 Instruction *Inst = &*Iter;
1135 if (Inst->getOpcode() != Instruction::Add)
1138 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
1139 if (!Inc || !Inc->isOne())
1142 PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry);
1146 // Check if the result of the instruction is live of the loop.
1147 bool LiveOutLoop = false;
1148 for (User *U : Inst->users()) {
1149 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
1166 // step 5: check if the precondition is in this form:
1167 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
1169 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1170 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
1171 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
1174 CntInst = CountInst;
1182 /// Return true if the idiom is detected in the loop.
1185 /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ)
1186 /// or nullptr if there is no such.
1187 /// 2) \p CntPhi is set to the corresponding phi node
1188 /// or nullptr if there is no such.
1189 /// 3) \p Var is set to the value whose CTLZ could be used.
1190 /// 4) \p DefX is set to the instruction calculating Loop exit condition.
1192 /// The core idiom we are trying to detect is:
1195 /// goto loop-exit // the precondition of the loop
1196 /// cnt0 = init-val;
1198 /// x = phi (x0, x.next); //PhiX
1199 /// cnt = phi(cnt0, cnt.next);
1201 /// cnt.next = cnt + 1;
1203 /// x.next = x >> 1; // DefX
1205 /// } while(x.next != 0);
1209 static bool detectCTLZIdiom(Loop *CurLoop, PHINode *&PhiX,
1210 Instruction *&CntInst, PHINode *&CntPhi,
1211 Instruction *&DefX) {
1212 BasicBlock *LoopEntry;
1213 Value *VarX = nullptr;
1219 LoopEntry = *(CurLoop->block_begin());
1221 // step 1: Check if the loop-back branch is in desirable form.
1222 if (Value *T = matchCondition(
1223 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1224 DefX = dyn_cast<Instruction>(T);
1228 // step 2: detect instructions corresponding to "x.next = x >> 1"
1229 if (!DefX || DefX->getOpcode() != Instruction::AShr)
1231 if (ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1)))
1232 if (!Shft || !Shft->isOne())
1234 VarX = DefX->getOperand(0);
1236 // step 3: Check the recurrence of variable X
1237 PhiX = getRecurrenceVar(VarX, DefX, LoopEntry);
1241 // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1
1242 // TODO: We can skip the step. If loop trip count is known (CTLZ),
1243 // then all uses of "cnt.next" could be optimized to the trip count
1244 // plus "cnt0". Currently it is not optimized.
1245 // This step could be used to detect POPCNT instruction:
1246 // cnt.next = cnt + (x.next & 1)
1247 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
1248 IterE = LoopEntry->end();
1249 Iter != IterE; Iter++) {
1250 Instruction *Inst = &*Iter;
1251 if (Inst->getOpcode() != Instruction::Add)
1254 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
1255 if (!Inc || !Inc->isOne())
1258 PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry);
1272 /// Recognize CTLZ idiom in a non-countable loop and convert the loop
1273 /// to countable (with CTLZ trip count).
1274 /// If CTLZ inserted as a new trip count returns true; otherwise, returns false.
1275 bool LoopIdiomRecognize::recognizeAndInsertCTLZ() {
1276 // Give up if the loop has multiple blocks or multiple backedges.
1277 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1280 Instruction *CntInst, *DefX;
1281 PHINode *CntPhi, *PhiX;
1282 if (!detectCTLZIdiom(CurLoop, PhiX, CntInst, CntPhi, DefX))
1285 bool IsCntPhiUsedOutsideLoop = false;
1286 for (User *U : CntPhi->users())
1287 if (!CurLoop->contains(dyn_cast<Instruction>(U))) {
1288 IsCntPhiUsedOutsideLoop = true;
1291 bool IsCntInstUsedOutsideLoop = false;
1292 for (User *U : CntInst->users())
1293 if (!CurLoop->contains(dyn_cast<Instruction>(U))) {
1294 IsCntInstUsedOutsideLoop = true;
1297 // If both CntInst and CntPhi are used outside the loop the profitability
1299 if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop)
1302 // For some CPUs result of CTLZ(X) intrinsic is undefined
1303 // when X is 0. If we can not guarantee X != 0, we need to check this
1305 bool ZeroCheck = false;
1306 // It is safe to assume Preheader exist as it was checked in
1307 // parent function RunOnLoop.
1308 BasicBlock *PH = CurLoop->getLoopPreheader();
1309 Value *InitX = PhiX->getIncomingValueForBlock(PH);
1310 // If we check X != 0 before entering the loop we don't need a zero
1311 // check in CTLZ intrinsic, but only if Cnt Phi is not used outside of the
1312 // loop (if it is used we count CTLZ(X >> 1)).
1313 if (!IsCntPhiUsedOutsideLoop)
1314 if (BasicBlock *PreCondBB = PH->getSinglePredecessor())
1315 if (BranchInst *PreCondBr =
1316 dyn_cast<BranchInst>(PreCondBB->getTerminator())) {
1317 if (matchCondition(PreCondBr, PH) == InitX)
1321 // Check if CTLZ intrinsic is profitable. Assume it is always profitable
1322 // if we delete the loop (the loop has only 6 instructions):
1323 // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ]
1324 // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ]
1325 // %shr = ashr %n.addr.0, 1
1326 // %tobool = icmp eq %shr, 0
1327 // %inc = add nsw %i.0, 1
1330 IRBuilder<> Builder(PH->getTerminator());
1331 SmallVector<const Value *, 2> Ops =
1332 {InitX, ZeroCheck ? Builder.getTrue() : Builder.getFalse()};
1333 ArrayRef<const Value *> Args(Ops);
1334 if (CurLoop->getHeader()->size() != 6 &&
1335 TTI->getIntrinsicCost(Intrinsic::ctlz, InitX->getType(), Args) >
1336 TargetTransformInfo::TCC_Basic)
1339 const DebugLoc DL = DefX->getDebugLoc();
1340 transformLoopToCountable(PH, CntInst, CntPhi, InitX, DL, ZeroCheck,
1341 IsCntPhiUsedOutsideLoop);
1345 /// Recognizes a population count idiom in a non-countable loop.
1347 /// If detected, transforms the relevant code to issue the popcount intrinsic
1348 /// function call, and returns true; otherwise, returns false.
1349 bool LoopIdiomRecognize::recognizePopcount() {
1350 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
1353 // Counting population are usually conducted by few arithmetic instructions.
1354 // Such instructions can be easily "absorbed" by vacant slots in a
1355 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
1356 // in a compact loop.
1358 // Give up if the loop has multiple blocks or multiple backedges.
1359 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1362 BasicBlock *LoopBody = *(CurLoop->block_begin());
1363 if (LoopBody->size() >= 20) {
1364 // The loop is too big, bail out.
1368 // It should have a preheader containing nothing but an unconditional branch.
1369 BasicBlock *PH = CurLoop->getLoopPreheader();
1370 if (!PH || &PH->front() != PH->getTerminator())
1372 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
1373 if (!EntryBI || EntryBI->isConditional())
1376 // It should have a precondition block where the generated popcount instrinsic
1377 // function can be inserted.
1378 auto *PreCondBB = PH->getSinglePredecessor();
1381 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1382 if (!PreCondBI || PreCondBI->isUnconditional())
1385 Instruction *CntInst;
1388 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
1391 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
1395 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
1396 const DebugLoc &DL) {
1397 Value *Ops[] = {Val};
1398 Type *Tys[] = {Val->getType()};
1400 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
1401 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
1402 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
1403 CI->setDebugLoc(DL);
1408 static CallInst *createCTLZIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
1409 const DebugLoc &DL, bool ZeroCheck) {
1410 Value *Ops[] = {Val, ZeroCheck ? IRBuilder.getTrue() : IRBuilder.getFalse()};
1411 Type *Tys[] = {Val->getType()};
1413 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
1414 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctlz, Tys);
1415 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
1416 CI->setDebugLoc(DL);
1421 /// Transform the following loop:
1423 /// CntPhi = PHI [Cnt0, CntInst]
1424 /// PhiX = PHI [InitX, DefX]
1425 /// CntInst = CntPhi + 1
1426 /// DefX = PhiX >> 1
1428 /// Br: loop if (DefX != 0)
1429 /// Use(CntPhi) or Use(CntInst)
1432 /// If CntPhi used outside the loop:
1433 /// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1)
1434 /// Count = CountPrev + 1
1436 /// Count = BitWidth(InitX) - CTLZ(InitX)
1438 /// CntPhi = PHI [Cnt0, CntInst]
1439 /// PhiX = PHI [InitX, DefX]
1440 /// PhiCount = PHI [Count, Dec]
1441 /// CntInst = CntPhi + 1
1442 /// DefX = PhiX >> 1
1443 /// Dec = PhiCount - 1
1445 /// Br: loop if (Dec != 0)
1446 /// Use(CountPrev + Cnt0) // Use(CntPhi)
1448 /// Use(Count + Cnt0) // Use(CntInst)
1450 /// If LOOP_BODY is empty the loop will be deleted.
1451 /// If CntInst and DefX are not used in LOOP_BODY they will be removed.
1452 void LoopIdiomRecognize::transformLoopToCountable(
1453 BasicBlock *Preheader, Instruction *CntInst, PHINode *CntPhi, Value *InitX,
1454 const DebugLoc DL, bool ZeroCheck, bool IsCntPhiUsedOutsideLoop) {
1455 BranchInst *PreheaderBr = dyn_cast<BranchInst>(Preheader->getTerminator());
1457 // Step 1: Insert the CTLZ instruction at the end of the preheader block
1458 // Count = BitWidth - CTLZ(InitX);
1459 // If there are uses of CntPhi create:
1460 // CountPrev = BitWidth - CTLZ(InitX >> 1);
1461 IRBuilder<> Builder(PreheaderBr);
1462 Builder.SetCurrentDebugLocation(DL);
1463 Value *CTLZ, *Count, *CountPrev, *NewCount, *InitXNext;
1465 if (IsCntPhiUsedOutsideLoop)
1466 InitXNext = Builder.CreateAShr(InitX,
1467 ConstantInt::get(InitX->getType(), 1));
1470 CTLZ = createCTLZIntrinsic(Builder, InitXNext, DL, ZeroCheck);
1471 Count = Builder.CreateSub(
1472 ConstantInt::get(CTLZ->getType(),
1473 CTLZ->getType()->getIntegerBitWidth()),
1475 if (IsCntPhiUsedOutsideLoop) {
1477 Count = Builder.CreateAdd(
1479 ConstantInt::get(CountPrev->getType(), 1));
1481 if (IsCntPhiUsedOutsideLoop)
1482 NewCount = Builder.CreateZExtOrTrunc(CountPrev,
1483 cast<IntegerType>(CntInst->getType()));
1485 NewCount = Builder.CreateZExtOrTrunc(Count,
1486 cast<IntegerType>(CntInst->getType()));
1488 // If the CTLZ counter's initial value is not zero, insert Add Inst.
1489 Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader);
1490 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
1491 if (!InitConst || !InitConst->isZero())
1492 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
1494 // Step 2: Insert new IV and loop condition:
1497 // PhiCount = PHI [Count, Dec]
1499 // Dec = PhiCount - 1
1501 // Br: loop if (Dec != 0)
1502 BasicBlock *Body = *(CurLoop->block_begin());
1503 auto *LbBr = dyn_cast<BranchInst>(Body->getTerminator());
1504 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
1505 Type *Ty = Count->getType();
1507 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
1509 Builder.SetInsertPoint(LbCond);
1510 Instruction *TcDec = cast<Instruction>(
1511 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
1512 "tcdec", false, true));
1514 TcPhi->addIncoming(Count, Preheader);
1515 TcPhi->addIncoming(TcDec, Body);
1517 CmpInst::Predicate Pred =
1518 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ;
1519 LbCond->setPredicate(Pred);
1520 LbCond->setOperand(0, TcDec);
1521 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
1523 // Step 3: All the references to the original counter outside
1524 // the loop are replaced with the NewCount -- the value returned from
1525 // __builtin_ctlz(x).
1526 if (IsCntPhiUsedOutsideLoop)
1527 CntPhi->replaceUsesOutsideBlock(NewCount, Body);
1529 CntInst->replaceUsesOutsideBlock(NewCount, Body);
1531 // step 4: Forget the "non-computable" trip-count SCEV associated with the
1532 // loop. The loop would otherwise not be deleted even if it becomes empty.
1533 SE->forgetLoop(CurLoop);
1536 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
1537 Instruction *CntInst,
1538 PHINode *CntPhi, Value *Var) {
1539 BasicBlock *PreHead = CurLoop->getLoopPreheader();
1540 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1541 const DebugLoc DL = CntInst->getDebugLoc();
1543 // Assuming before transformation, the loop is following:
1544 // if (x) // the precondition
1545 // do { cnt++; x &= x - 1; } while(x);
1547 // Step 1: Insert the ctpop instruction at the end of the precondition block
1548 IRBuilder<> Builder(PreCondBr);
1549 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
1551 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
1552 NewCount = PopCntZext =
1553 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
1555 if (NewCount != PopCnt)
1556 (cast<Instruction>(NewCount))->setDebugLoc(DL);
1558 // TripCnt is exactly the number of iterations the loop has
1561 // If the population counter's initial value is not zero, insert Add Inst.
1562 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
1563 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
1564 if (!InitConst || !InitConst->isZero()) {
1565 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
1566 (cast<Instruction>(NewCount))->setDebugLoc(DL);
1570 // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
1571 // "if (NewCount == 0) loop-exit". Without this change, the intrinsic
1572 // function would be partial dead code, and downstream passes will drag
1573 // it back from the precondition block to the preheader.
1575 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
1577 Value *Opnd0 = PopCntZext;
1578 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
1579 if (PreCond->getOperand(0) != Var)
1580 std::swap(Opnd0, Opnd1);
1582 ICmpInst *NewPreCond = cast<ICmpInst>(
1583 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
1584 PreCondBr->setCondition(NewPreCond);
1586 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
1589 // Step 3: Note that the population count is exactly the trip count of the
1590 // loop in question, which enable us to to convert the loop from noncountable
1591 // loop into a countable one. The benefit is twofold:
1593 // - If the loop only counts population, the entire loop becomes dead after
1594 // the transformation. It is a lot easier to prove a countable loop dead
1595 // than to prove a noncountable one. (In some C dialects, an infinite loop
1596 // isn't dead even if it computes nothing useful. In general, DCE needs
1597 // to prove a noncountable loop finite before safely delete it.)
1599 // - If the loop also performs something else, it remains alive.
1600 // Since it is transformed to countable form, it can be aggressively
1601 // optimized by some optimizations which are in general not applicable
1602 // to a noncountable loop.
1604 // After this step, this loop (conceptually) would look like following:
1605 // newcnt = __builtin_ctpop(x);
1608 // do { cnt++; x &= x-1; t--) } while (t > 0);
1609 BasicBlock *Body = *(CurLoop->block_begin());
1611 auto *LbBr = dyn_cast<BranchInst>(Body->getTerminator());
1612 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
1613 Type *Ty = TripCnt->getType();
1615 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
1617 Builder.SetInsertPoint(LbCond);
1618 Instruction *TcDec = cast<Instruction>(
1619 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
1620 "tcdec", false, true));
1622 TcPhi->addIncoming(TripCnt, PreHead);
1623 TcPhi->addIncoming(TcDec, Body);
1625 CmpInst::Predicate Pred =
1626 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
1627 LbCond->setPredicate(Pred);
1628 LbCond->setOperand(0, TcDec);
1629 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
1632 // Step 4: All the references to the original population counter outside
1633 // the loop are replaced with the NewCount -- the value returned from
1634 // __builtin_ctpop().
1635 CntInst->replaceUsesOutsideBlock(NewCount, Body);
1637 // step 5: Forget the "non-computable" trip-count SCEV associated with the
1638 // loop. The loop would otherwise not be deleted even if it becomes empty.
1639 SE->forgetLoop(CurLoop);