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;
114 /// \name Countable Loop Idiom Handling
117 bool runOnCountableLoop();
118 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
119 SmallVectorImpl<BasicBlock *> &ExitBlocks);
121 void collectStores(BasicBlock *BB);
122 bool isLegalStore(StoreInst *SI, bool &ForMemset, bool &ForMemsetPattern,
124 bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount,
126 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
128 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
129 unsigned StoreAlignment, Value *StoredVal,
130 Instruction *TheStore,
131 SmallPtrSetImpl<Instruction *> &Stores,
132 const SCEVAddRecExpr *Ev, const SCEV *BECount,
133 bool NegStride, bool IsLoopMemset = false);
134 bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount);
135 bool avoidLIRForMultiBlockLoop(bool IsMemset = false,
136 bool IsLoopMemset = false);
139 /// \name Noncountable Loop Idiom Handling
142 bool runOnNoncountableLoop();
144 bool recognizePopcount();
145 void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
146 PHINode *CntPhi, Value *Var);
147 bool recognizeAndInsertCTLZ();
148 void transformLoopToCountable(BasicBlock *PreCondBB, Instruction *CntInst,
149 PHINode *CntPhi, Value *Var, const DebugLoc DL,
150 bool ZeroCheck, bool IsCntPhiUsedOutsideLoop);
155 class LoopIdiomRecognizeLegacyPass : public LoopPass {
158 explicit LoopIdiomRecognizeLegacyPass() : LoopPass(ID) {
159 initializeLoopIdiomRecognizeLegacyPassPass(
160 *PassRegistry::getPassRegistry());
163 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
167 AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
168 DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
169 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
170 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
171 TargetLibraryInfo *TLI =
172 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
173 const TargetTransformInfo *TTI =
174 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
175 *L->getHeader()->getParent());
176 const DataLayout *DL = &L->getHeader()->getModule()->getDataLayout();
178 LoopIdiomRecognize LIR(AA, DT, LI, SE, TLI, TTI, DL);
179 return LIR.runOnLoop(L);
182 /// This transformation requires natural loop information & requires that
183 /// loop preheaders be inserted into the CFG.
185 void getAnalysisUsage(AnalysisUsage &AU) const override {
186 AU.addRequired<TargetLibraryInfoWrapperPass>();
187 AU.addRequired<TargetTransformInfoWrapperPass>();
188 getLoopAnalysisUsage(AU);
191 } // End anonymous namespace.
193 PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM,
194 LoopStandardAnalysisResults &AR,
196 const auto *DL = &L.getHeader()->getModule()->getDataLayout();
198 LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI, DL);
199 if (!LIR.runOnLoop(&L))
200 return PreservedAnalyses::all();
202 return getLoopPassPreservedAnalyses();
205 char LoopIdiomRecognizeLegacyPass::ID = 0;
206 INITIALIZE_PASS_BEGIN(LoopIdiomRecognizeLegacyPass, "loop-idiom",
207 "Recognize loop idioms", false, false)
208 INITIALIZE_PASS_DEPENDENCY(LoopPass)
209 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
210 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
211 INITIALIZE_PASS_END(LoopIdiomRecognizeLegacyPass, "loop-idiom",
212 "Recognize loop idioms", false, false)
214 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognizeLegacyPass(); }
216 static void deleteDeadInstruction(Instruction *I) {
217 I->replaceAllUsesWith(UndefValue::get(I->getType()));
218 I->eraseFromParent();
221 //===----------------------------------------------------------------------===//
223 // Implementation of LoopIdiomRecognize
225 //===----------------------------------------------------------------------===//
227 bool LoopIdiomRecognize::runOnLoop(Loop *L) {
229 // If the loop could not be converted to canonical form, it must have an
230 // indirectbr in it, just give up.
231 if (!L->getLoopPreheader())
234 // Disable loop idiom recognition if the function's name is a common idiom.
235 StringRef Name = L->getHeader()->getParent()->getName();
236 if (Name == "memset" || Name == "memcpy")
239 // Determine if code size heuristics need to be applied.
240 ApplyCodeSizeHeuristics =
241 L->getHeader()->getParent()->optForSize() && UseLIRCodeSizeHeurs;
243 HasMemset = TLI->has(LibFunc_memset);
244 HasMemsetPattern = TLI->has(LibFunc_memset_pattern16);
245 HasMemcpy = TLI->has(LibFunc_memcpy);
247 if (HasMemset || HasMemsetPattern || HasMemcpy)
248 if (SE->hasLoopInvariantBackedgeTakenCount(L))
249 return runOnCountableLoop();
251 return runOnNoncountableLoop();
254 bool LoopIdiomRecognize::runOnCountableLoop() {
255 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
256 assert(!isa<SCEVCouldNotCompute>(BECount) &&
257 "runOnCountableLoop() called on a loop without a predictable"
258 "backedge-taken count");
260 // If this loop executes exactly one time, then it should be peeled, not
261 // optimized by this pass.
262 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
263 if (BECst->getAPInt() == 0)
266 SmallVector<BasicBlock *, 8> ExitBlocks;
267 CurLoop->getUniqueExitBlocks(ExitBlocks);
269 DEBUG(dbgs() << "loop-idiom Scanning: F["
270 << CurLoop->getHeader()->getParent()->getName() << "] Loop %"
271 << CurLoop->getHeader()->getName() << "\n");
273 bool MadeChange = false;
275 // The following transforms hoist stores/memsets into the loop pre-header.
276 // Give up if the loop has instructions may throw.
277 LoopSafetyInfo SafetyInfo;
278 computeLoopSafetyInfo(&SafetyInfo, CurLoop);
279 if (SafetyInfo.MayThrow)
282 // Scan all the blocks in the loop that are not in subloops.
283 for (auto *BB : CurLoop->getBlocks()) {
284 // Ignore blocks in subloops.
285 if (LI->getLoopFor(BB) != CurLoop)
288 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
293 static unsigned getStoreSizeInBytes(StoreInst *SI, const DataLayout *DL) {
294 uint64_t SizeInBits = DL->getTypeSizeInBits(SI->getValueOperand()->getType());
295 assert(((SizeInBits & 7) || (SizeInBits >> 32) == 0) &&
296 "Don't overflow unsigned.");
297 return (unsigned)SizeInBits >> 3;
300 static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) {
301 const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1));
302 return ConstStride->getAPInt();
305 /// getMemSetPatternValue - If a strided store of the specified value is safe to
306 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
307 /// be passed in. Otherwise, return null.
309 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
310 /// just replicate their input array and then pass on to memset_pattern16.
311 static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) {
312 // If the value isn't a constant, we can't promote it to being in a constant
313 // array. We could theoretically do a store to an alloca or something, but
314 // that doesn't seem worthwhile.
315 Constant *C = dyn_cast<Constant>(V);
319 // Only handle simple values that are a power of two bytes in size.
320 uint64_t Size = DL->getTypeSizeInBits(V->getType());
321 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
324 // Don't care enough about darwin/ppc to implement this.
325 if (DL->isBigEndian())
328 // Convert to size in bytes.
331 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
332 // if the top and bottom are the same (e.g. for vectors and large integers).
336 // If the constant is exactly 16 bytes, just use it.
340 // Otherwise, we'll use an array of the constants.
341 unsigned ArraySize = 16 / Size;
342 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
343 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
346 bool LoopIdiomRecognize::isLegalStore(StoreInst *SI, bool &ForMemset,
347 bool &ForMemsetPattern, bool &ForMemcpy) {
348 // Don't touch volatile stores.
352 // Don't convert stores of non-integral pointer types to memsets (which stores
354 if (DL->isNonIntegralPointerType(SI->getValueOperand()->getType()))
357 // Avoid merging nontemporal stores.
358 if (SI->getMetadata(LLVMContext::MD_nontemporal))
361 Value *StoredVal = SI->getValueOperand();
362 Value *StorePtr = SI->getPointerOperand();
364 // Reject stores that are so large that they overflow an unsigned.
365 uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
366 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
369 // See if the pointer expression is an AddRec like {base,+,1} on the current
370 // loop, which indicates a strided store. If we have something else, it's a
371 // random store we can't handle.
372 const SCEVAddRecExpr *StoreEv =
373 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
374 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
377 // Check to see if we have a constant stride.
378 if (!isa<SCEVConstant>(StoreEv->getOperand(1)))
381 // See if the store can be turned into a memset.
383 // If the stored value is a byte-wise value (like i32 -1), then it may be
384 // turned into a memset of i8 -1, assuming that all the consecutive bytes
385 // are stored. A store of i32 0x01020304 can never be turned into a memset,
386 // but it can be turned into memset_pattern if the target supports it.
387 Value *SplatValue = isBytewiseValue(StoredVal);
388 Constant *PatternValue = nullptr;
390 // If we're allowed to form a memset, and the stored value would be
391 // acceptable for memset, use it.
392 if (HasMemset && SplatValue &&
393 // Verify that the stored value is loop invariant. If not, we can't
394 // promote the memset.
395 CurLoop->isLoopInvariant(SplatValue)) {
396 // It looks like we can use SplatValue.
399 } else if (HasMemsetPattern &&
400 // Don't create memset_pattern16s with address spaces.
401 StorePtr->getType()->getPointerAddressSpace() == 0 &&
402 (PatternValue = getMemSetPatternValue(StoredVal, DL))) {
403 // It looks like we can use PatternValue!
404 ForMemsetPattern = true;
408 // Otherwise, see if the store can be turned into a memcpy.
410 // Check to see if the stride matches the size of the store. If so, then we
411 // know that every byte is touched in the loop.
412 APInt Stride = getStoreStride(StoreEv);
413 unsigned StoreSize = getStoreSizeInBytes(SI, DL);
414 if (StoreSize != Stride && StoreSize != -Stride)
417 // The store must be feeding a non-volatile load.
418 LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand());
419 if (!LI || !LI->isSimple())
422 // See if the pointer expression is an AddRec like {base,+,1} on the current
423 // loop, which indicates a strided load. If we have something else, it's a
424 // random load we can't handle.
425 const SCEVAddRecExpr *LoadEv =
426 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
427 if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
430 // The store and load must share the same stride.
431 if (StoreEv->getOperand(1) != LoadEv->getOperand(1))
434 // Success. This store can be converted into a memcpy.
438 // This store can't be transformed into a memset/memcpy.
442 void LoopIdiomRecognize::collectStores(BasicBlock *BB) {
443 StoreRefsForMemset.clear();
444 StoreRefsForMemsetPattern.clear();
445 StoreRefsForMemcpy.clear();
446 for (Instruction &I : *BB) {
447 StoreInst *SI = dyn_cast<StoreInst>(&I);
451 bool ForMemset = false;
452 bool ForMemsetPattern = false;
453 bool ForMemcpy = false;
454 // Make sure this is a strided store with a constant stride.
455 if (!isLegalStore(SI, ForMemset, ForMemsetPattern, ForMemcpy))
458 // Save the store locations.
460 // Find the base pointer.
461 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), *DL);
462 StoreRefsForMemset[Ptr].push_back(SI);
463 } else if (ForMemsetPattern) {
464 // Find the base pointer.
465 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), *DL);
466 StoreRefsForMemsetPattern[Ptr].push_back(SI);
467 } else if (ForMemcpy)
468 StoreRefsForMemcpy.push_back(SI);
472 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
473 /// with the specified backedge count. This block is known to be in the current
474 /// loop and not in any subloops.
475 bool LoopIdiomRecognize::runOnLoopBlock(
476 BasicBlock *BB, const SCEV *BECount,
477 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
478 // We can only promote stores in this block if they are unconditionally
479 // executed in the loop. For a block to be unconditionally executed, it has
480 // to dominate all the exit blocks of the loop. Verify this now.
481 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
482 if (!DT->dominates(BB, ExitBlocks[i]))
485 bool MadeChange = false;
486 // Look for store instructions, which may be optimized to memset/memcpy.
489 // Look for a single store or sets of stores with a common base, which can be
490 // optimized into a memset (memset_pattern). The latter most commonly happens
491 // with structs and handunrolled loops.
492 for (auto &SL : StoreRefsForMemset)
493 MadeChange |= processLoopStores(SL.second, BECount, true);
495 for (auto &SL : StoreRefsForMemsetPattern)
496 MadeChange |= processLoopStores(SL.second, BECount, false);
498 // Optimize the store into a memcpy, if it feeds an similarly strided load.
499 for (auto &SI : StoreRefsForMemcpy)
500 MadeChange |= processLoopStoreOfLoopLoad(SI, BECount);
502 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
503 Instruction *Inst = &*I++;
504 // Look for memset instructions, which may be optimized to a larger memset.
505 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
506 WeakTrackingVH InstPtr(&*I);
507 if (!processLoopMemSet(MSI, BECount))
511 // If processing the memset invalidated our iterator, start over from the
522 /// processLoopStores - See if this store(s) can be promoted to a memset.
523 bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL,
526 // Try to find consecutive stores that can be transformed into memsets.
527 SetVector<StoreInst *> Heads, Tails;
528 SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain;
530 // Do a quadratic search on all of the given stores and find
531 // all of the pairs of stores that follow each other.
532 SmallVector<unsigned, 16> IndexQueue;
533 for (unsigned i = 0, e = SL.size(); i < e; ++i) {
534 assert(SL[i]->isSimple() && "Expected only non-volatile stores.");
536 Value *FirstStoredVal = SL[i]->getValueOperand();
537 Value *FirstStorePtr = SL[i]->getPointerOperand();
538 const SCEVAddRecExpr *FirstStoreEv =
539 cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr));
540 APInt FirstStride = getStoreStride(FirstStoreEv);
541 unsigned FirstStoreSize = getStoreSizeInBytes(SL[i], DL);
543 // See if we can optimize just this store in isolation.
544 if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) {
549 Value *FirstSplatValue = nullptr;
550 Constant *FirstPatternValue = nullptr;
553 FirstSplatValue = isBytewiseValue(FirstStoredVal);
555 FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL);
557 assert((FirstSplatValue || FirstPatternValue) &&
558 "Expected either splat value or pattern value.");
561 // If a store has multiple consecutive store candidates, search Stores
562 // array according to the sequence: from i+1 to e, then from i-1 to 0.
563 // This is because usually pairing with immediate succeeding or preceding
564 // candidate create the best chance to find memset opportunity.
566 for (j = i + 1; j < e; ++j)
567 IndexQueue.push_back(j);
568 for (j = i; j > 0; --j)
569 IndexQueue.push_back(j - 1);
571 for (auto &k : IndexQueue) {
572 assert(SL[k]->isSimple() && "Expected only non-volatile stores.");
573 Value *SecondStorePtr = SL[k]->getPointerOperand();
574 const SCEVAddRecExpr *SecondStoreEv =
575 cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr));
576 APInt SecondStride = getStoreStride(SecondStoreEv);
578 if (FirstStride != SecondStride)
581 Value *SecondStoredVal = SL[k]->getValueOperand();
582 Value *SecondSplatValue = nullptr;
583 Constant *SecondPatternValue = nullptr;
586 SecondSplatValue = isBytewiseValue(SecondStoredVal);
588 SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL);
590 assert((SecondSplatValue || SecondPatternValue) &&
591 "Expected either splat value or pattern value.");
593 if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) {
595 if (FirstSplatValue != SecondSplatValue)
598 if (FirstPatternValue != SecondPatternValue)
603 ConsecutiveChain[SL[i]] = SL[k];
609 // We may run into multiple chains that merge into a single chain. We mark the
610 // stores that we transformed so that we don't visit the same store twice.
611 SmallPtrSet<Value *, 16> TransformedStores;
612 bool Changed = false;
614 // For stores that start but don't end a link in the chain:
615 for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end();
617 if (Tails.count(*it))
620 // We found a store instr that starts a chain. Now follow the chain and try
622 SmallPtrSet<Instruction *, 8> AdjacentStores;
625 StoreInst *HeadStore = I;
626 unsigned StoreSize = 0;
628 // Collect the chain into a list.
629 while (Tails.count(I) || Heads.count(I)) {
630 if (TransformedStores.count(I))
632 AdjacentStores.insert(I);
634 StoreSize += getStoreSizeInBytes(I, DL);
635 // Move to the next value in the chain.
636 I = ConsecutiveChain[I];
639 Value *StoredVal = HeadStore->getValueOperand();
640 Value *StorePtr = HeadStore->getPointerOperand();
641 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
642 APInt Stride = getStoreStride(StoreEv);
644 // Check to see if the stride matches the size of the stores. If so, then
645 // we know that every byte is touched in the loop.
646 if (StoreSize != Stride && StoreSize != -Stride)
649 bool NegStride = StoreSize == -Stride;
651 if (processLoopStridedStore(StorePtr, StoreSize, HeadStore->getAlignment(),
652 StoredVal, HeadStore, AdjacentStores, StoreEv,
653 BECount, NegStride)) {
654 TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end());
662 /// processLoopMemSet - See if this memset can be promoted to a large memset.
663 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
664 const SCEV *BECount) {
665 // We can only handle non-volatile memsets with a constant size.
666 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength()))
669 // If we're not allowed to hack on memset, we fail.
673 Value *Pointer = MSI->getDest();
675 // See if the pointer expression is an AddRec like {base,+,1} on the current
676 // loop, which indicates a strided store. If we have something else, it's a
677 // random store we can't handle.
678 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
679 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
682 // Reject memsets that are so large that they overflow an unsigned.
683 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
684 if ((SizeInBytes >> 32) != 0)
687 // Check to see if the stride matches the size of the memset. If so, then we
688 // know that every byte is touched in the loop.
689 const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
693 APInt Stride = ConstStride->getAPInt();
694 if (SizeInBytes != Stride && SizeInBytes != -Stride)
697 // Verify that the memset value is loop invariant. If not, we can't promote
699 Value *SplatValue = MSI->getValue();
700 if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue))
703 SmallPtrSet<Instruction *, 1> MSIs;
705 bool NegStride = SizeInBytes == -Stride;
706 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
707 MSI->getAlignment(), SplatValue, MSI, MSIs, Ev,
708 BECount, NegStride, /*IsLoopMemset=*/true);
711 /// mayLoopAccessLocation - Return true if the specified loop might access the
712 /// specified pointer location, which is a loop-strided access. The 'Access'
713 /// argument specifies what the verboten forms of access are (read or write).
715 mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
716 const SCEV *BECount, unsigned StoreSize,
718 SmallPtrSetImpl<Instruction *> &IgnoredStores) {
719 // Get the location that may be stored across the loop. Since the access is
720 // strided positively through memory, we say that the modified location starts
721 // at the pointer and has infinite size.
722 uint64_t AccessSize = MemoryLocation::UnknownSize;
724 // If the loop iterates a fixed number of times, we can refine the access size
725 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
726 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
727 AccessSize = (BECst->getValue()->getZExtValue() + 1) * StoreSize;
729 // TODO: For this to be really effective, we have to dive into the pointer
730 // operand in the store. Store to &A[i] of 100 will always return may alias
731 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
732 // which will then no-alias a store to &A[100].
733 MemoryLocation StoreLoc(Ptr, AccessSize);
735 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
737 for (Instruction &I : **BI)
738 if (IgnoredStores.count(&I) == 0 &&
739 (AA.getModRefInfo(&I, StoreLoc) & Access))
745 // If we have a negative stride, Start refers to the end of the memory location
746 // we're trying to memset. Therefore, we need to recompute the base pointer,
747 // which is just Start - BECount*Size.
748 static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount,
749 Type *IntPtr, unsigned StoreSize,
750 ScalarEvolution *SE) {
751 const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr);
753 Index = SE->getMulExpr(Index, SE->getConstant(IntPtr, StoreSize),
755 return SE->getMinusSCEV(Start, Index);
758 /// processLoopStridedStore - We see a strided store of some value. If we can
759 /// transform this into a memset or memset_pattern in the loop preheader, do so.
760 bool LoopIdiomRecognize::processLoopStridedStore(
761 Value *DestPtr, unsigned StoreSize, unsigned StoreAlignment,
762 Value *StoredVal, Instruction *TheStore,
763 SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev,
764 const SCEV *BECount, bool NegStride, bool IsLoopMemset) {
765 Value *SplatValue = isBytewiseValue(StoredVal);
766 Constant *PatternValue = nullptr;
769 PatternValue = getMemSetPatternValue(StoredVal, DL);
771 assert((SplatValue || PatternValue) &&
772 "Expected either splat value or pattern value.");
774 // The trip count of the loop and the base pointer of the addrec SCEV is
775 // guaranteed to be loop invariant, which means that it should dominate the
776 // header. This allows us to insert code for it in the preheader.
777 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
778 BasicBlock *Preheader = CurLoop->getLoopPreheader();
779 IRBuilder<> Builder(Preheader->getTerminator());
780 SCEVExpander Expander(*SE, *DL, "loop-idiom");
782 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
783 Type *IntPtr = Builder.getIntPtrTy(*DL, DestAS);
785 const SCEV *Start = Ev->getStart();
786 // Handle negative strided loops.
788 Start = getStartForNegStride(Start, BECount, IntPtr, StoreSize, SE);
790 // TODO: ideally we should still be able to generate memset if SCEV expander
791 // is taught to generate the dependencies at the latest point.
792 if (!isSafeToExpand(Start, *SE))
795 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
796 // this into a memset in the loop preheader now if we want. However, this
797 // would be unsafe to do if there is anything else in the loop that may read
798 // or write to the aliased location. Check for any overlap by generating the
799 // base pointer and checking the region.
801 Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator());
802 if (mayLoopAccessLocation(BasePtr, MRI_ModRef, CurLoop, BECount, StoreSize,
805 // If we generated new code for the base pointer, clean up.
806 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
810 if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset))
813 // Okay, everything looks good, insert the memset.
815 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
816 // pointer size if it isn't already.
817 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtr);
819 const SCEV *NumBytesS =
820 SE->getAddExpr(BECount, SE->getOne(IntPtr), SCEV::FlagNUW);
821 if (StoreSize != 1) {
822 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
826 // TODO: ideally we should still be able to generate memset if SCEV expander
827 // is taught to generate the dependencies at the latest point.
828 if (!isSafeToExpand(NumBytesS, *SE))
832 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
837 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment);
839 // Everything is emitted in default address space
840 Type *Int8PtrTy = DestInt8PtrTy;
842 Module *M = TheStore->getModule();
844 M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(),
845 Int8PtrTy, Int8PtrTy, IntPtr);
846 inferLibFuncAttributes(*M->getFunction("memset_pattern16"), *TLI);
848 // Otherwise we should form a memset_pattern16. PatternValue is known to be
849 // an constant array of 16-bytes. Plop the value into a mergable global.
850 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
851 GlobalValue::PrivateLinkage,
852 PatternValue, ".memset_pattern");
853 GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these.
854 GV->setAlignment(16);
855 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
856 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
859 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
860 << " from store to: " << *Ev << " at: " << *TheStore << "\n");
861 NewCall->setDebugLoc(TheStore->getDebugLoc());
863 // Okay, the memset has been formed. Zap the original store and anything that
865 for (auto *I : Stores)
866 deleteDeadInstruction(I);
871 /// If the stored value is a strided load in the same loop with the same stride
872 /// this may be transformable into a memcpy. This kicks in for stuff like
873 /// for (i) A[i] = B[i];
874 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI,
875 const SCEV *BECount) {
876 assert(SI->isSimple() && "Expected only non-volatile stores.");
878 Value *StorePtr = SI->getPointerOperand();
879 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
880 APInt Stride = getStoreStride(StoreEv);
881 unsigned StoreSize = getStoreSizeInBytes(SI, DL);
882 bool NegStride = StoreSize == -Stride;
884 // The store must be feeding a non-volatile load.
885 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
886 assert(LI->isSimple() && "Expected only non-volatile stores.");
888 // See if the pointer expression is an AddRec like {base,+,1} on the current
889 // loop, which indicates a strided load. If we have something else, it's a
890 // random load we can't handle.
891 const SCEVAddRecExpr *LoadEv =
892 cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
894 // The trip count of the loop and the base pointer of the addrec SCEV is
895 // guaranteed to be loop invariant, which means that it should dominate the
896 // header. This allows us to insert code for it in the preheader.
897 BasicBlock *Preheader = CurLoop->getLoopPreheader();
898 IRBuilder<> Builder(Preheader->getTerminator());
899 SCEVExpander Expander(*SE, *DL, "loop-idiom");
901 const SCEV *StrStart = StoreEv->getStart();
902 unsigned StrAS = SI->getPointerAddressSpace();
903 Type *IntPtrTy = Builder.getIntPtrTy(*DL, StrAS);
905 // Handle negative strided loops.
907 StrStart = getStartForNegStride(StrStart, BECount, IntPtrTy, StoreSize, SE);
909 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
910 // this into a memcpy in the loop preheader now if we want. However, this
911 // would be unsafe to do if there is anything else in the loop that may read
912 // or write the memory region we're storing to. This includes the load that
913 // feeds the stores. Check for an alias by generating the base address and
914 // checking everything.
915 Value *StoreBasePtr = Expander.expandCodeFor(
916 StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator());
918 SmallPtrSet<Instruction *, 1> Stores;
920 if (mayLoopAccessLocation(StoreBasePtr, MRI_ModRef, CurLoop, BECount,
921 StoreSize, *AA, Stores)) {
923 // If we generated new code for the base pointer, clean up.
924 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
928 const SCEV *LdStart = LoadEv->getStart();
929 unsigned LdAS = LI->getPointerAddressSpace();
931 // Handle negative strided loops.
933 LdStart = getStartForNegStride(LdStart, BECount, IntPtrTy, StoreSize, SE);
935 // For a memcpy, we have to make sure that the input array is not being
936 // mutated by the loop.
937 Value *LoadBasePtr = Expander.expandCodeFor(
938 LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator());
940 if (mayLoopAccessLocation(LoadBasePtr, MRI_Mod, CurLoop, BECount, StoreSize,
943 // If we generated new code for the base pointer, clean up.
944 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
945 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
949 if (avoidLIRForMultiBlockLoop())
952 // Okay, everything is safe, we can transform this!
954 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
955 // pointer size if it isn't already.
956 BECount = SE->getTruncateOrZeroExtend(BECount, IntPtrTy);
958 const SCEV *NumBytesS =
959 SE->getAddExpr(BECount, SE->getOne(IntPtrTy), SCEV::FlagNUW);
961 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtrTy, StoreSize),
965 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
968 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
969 std::min(SI->getAlignment(), LI->getAlignment()));
970 NewCall->setDebugLoc(SI->getDebugLoc());
972 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
973 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
974 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
976 // Okay, the memcpy has been formed. Zap the original store and anything that
978 deleteDeadInstruction(SI);
983 // When compiling for codesize we avoid idiom recognition for a multi-block loop
984 // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop.
986 bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset,
988 if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) {
989 if (!CurLoop->getParentLoop() && (!IsMemset || !IsLoopMemset)) {
990 DEBUG(dbgs() << " " << CurLoop->getHeader()->getParent()->getName()
991 << " : LIR " << (IsMemset ? "Memset" : "Memcpy")
992 << " avoided: multi-block top-level loop\n");
1000 bool LoopIdiomRecognize::runOnNoncountableLoop() {
1001 return recognizePopcount() || recognizeAndInsertCTLZ();
1004 /// Check if the given conditional branch is based on the comparison between
1005 /// a variable and zero, and if the variable is non-zero, the control yields to
1006 /// the loop entry. If the branch matches the behavior, the variable involved
1007 /// in the comparison is returned. This function will be called to see if the
1008 /// precondition and postcondition of the loop are in desirable form.
1009 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry) {
1010 if (!BI || !BI->isConditional())
1013 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1017 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
1018 if (!CmpZero || !CmpZero->isZero())
1021 ICmpInst::Predicate Pred = Cond->getPredicate();
1022 if ((Pred == ICmpInst::ICMP_NE && BI->getSuccessor(0) == LoopEntry) ||
1023 (Pred == ICmpInst::ICMP_EQ && BI->getSuccessor(1) == LoopEntry))
1024 return Cond->getOperand(0);
1029 /// Return true iff the idiom is detected in the loop.
1032 /// 1) \p CntInst is set to the instruction counting the population bit.
1033 /// 2) \p CntPhi is set to the corresponding phi node.
1034 /// 3) \p Var is set to the value whose population bits are being counted.
1036 /// The core idiom we are trying to detect is:
1039 /// goto loop-exit // the precondition of the loop
1040 /// cnt0 = init-val;
1042 /// x1 = phi (x0, x2);
1043 /// cnt1 = phi(cnt0, cnt2);
1045 /// cnt2 = cnt1 + 1;
1047 /// x2 = x1 & (x1 - 1);
1049 /// } while(x != 0);
1053 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
1054 Instruction *&CntInst, PHINode *&CntPhi,
1056 // step 1: Check to see if the look-back branch match this pattern:
1057 // "if (a!=0) goto loop-entry".
1058 BasicBlock *LoopEntry;
1059 Instruction *DefX2, *CountInst;
1060 Value *VarX1, *VarX0;
1061 PHINode *PhiX, *CountPhi;
1063 DefX2 = CountInst = nullptr;
1064 VarX1 = VarX0 = nullptr;
1065 PhiX = CountPhi = nullptr;
1066 LoopEntry = *(CurLoop->block_begin());
1068 // step 1: Check if the loop-back branch is in desirable form.
1070 if (Value *T = matchCondition(
1071 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1072 DefX2 = dyn_cast<Instruction>(T);
1077 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
1079 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
1082 BinaryOperator *SubOneOp;
1084 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
1085 VarX1 = DefX2->getOperand(1);
1087 VarX1 = DefX2->getOperand(0);
1088 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
1093 Instruction *SubInst = cast<Instruction>(SubOneOp);
1094 ConstantInt *Dec = dyn_cast<ConstantInt>(SubInst->getOperand(1));
1096 !((SubInst->getOpcode() == Instruction::Sub && Dec->isOne()) ||
1097 (SubInst->getOpcode() == Instruction::Add &&
1098 Dec->isAllOnesValue()))) {
1103 // step 3: Check the recurrence of variable X
1105 PhiX = dyn_cast<PHINode>(VarX1);
1107 (PhiX->getOperand(0) != DefX2 && PhiX->getOperand(1) != DefX2)) {
1112 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
1114 CountInst = nullptr;
1115 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
1116 IterE = LoopEntry->end();
1117 Iter != IterE; Iter++) {
1118 Instruction *Inst = &*Iter;
1119 if (Inst->getOpcode() != Instruction::Add)
1122 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
1123 if (!Inc || !Inc->isOne())
1126 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
1127 if (!Phi || Phi->getParent() != LoopEntry)
1130 // Check if the result of the instruction is live of the loop.
1131 bool LiveOutLoop = false;
1132 for (User *U : Inst->users()) {
1133 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
1150 // step 5: check if the precondition is in this form:
1151 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
1153 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1154 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
1155 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
1158 CntInst = CountInst;
1166 /// Return true if the idiom is detected in the loop.
1169 /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ)
1170 /// or nullptr if there is no such.
1171 /// 2) \p CntPhi is set to the corresponding phi node
1172 /// or nullptr if there is no such.
1173 /// 3) \p Var is set to the value whose CTLZ could be used.
1174 /// 4) \p DefX is set to the instruction calculating Loop exit condition.
1176 /// The core idiom we are trying to detect is:
1179 /// goto loop-exit // the precondition of the loop
1180 /// cnt0 = init-val;
1182 /// x = phi (x0, x.next); //PhiX
1183 /// cnt = phi(cnt0, cnt.next);
1185 /// cnt.next = cnt + 1;
1187 /// x.next = x >> 1; // DefX
1189 /// } while(x.next != 0);
1193 static bool detectCTLZIdiom(Loop *CurLoop, PHINode *&PhiX,
1194 Instruction *&CntInst, PHINode *&CntPhi,
1195 Instruction *&DefX) {
1196 BasicBlock *LoopEntry;
1197 Value *VarX = nullptr;
1203 LoopEntry = *(CurLoop->block_begin());
1205 // step 1: Check if the loop-back branch is in desirable form.
1206 if (Value *T = matchCondition(
1207 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1208 DefX = dyn_cast<Instruction>(T);
1212 // step 2: detect instructions corresponding to "x.next = x >> 1"
1213 if (!DefX || DefX->getOpcode() != Instruction::AShr)
1215 if (ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1)))
1216 if (!Shft || !Shft->isOne())
1218 VarX = DefX->getOperand(0);
1220 // step 3: Check the recurrence of variable X
1221 PhiX = dyn_cast<PHINode>(VarX);
1222 if (!PhiX || (PhiX->getOperand(0) != DefX && PhiX->getOperand(1) != DefX))
1225 // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1
1226 // TODO: We can skip the step. If loop trip count is known (CTLZ),
1227 // then all uses of "cnt.next" could be optimized to the trip count
1228 // plus "cnt0". Currently it is not optimized.
1229 // This step could be used to detect POPCNT instruction:
1230 // cnt.next = cnt + (x.next & 1)
1231 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
1232 IterE = LoopEntry->end();
1233 Iter != IterE; Iter++) {
1234 Instruction *Inst = &*Iter;
1235 if (Inst->getOpcode() != Instruction::Add)
1238 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
1239 if (!Inc || !Inc->isOne())
1242 PHINode *Phi = dyn_cast<PHINode>(Inst->getOperand(0));
1243 if (!Phi || Phi->getParent() != LoopEntry)
1256 /// Recognize CTLZ idiom in a non-countable loop and convert the loop
1257 /// to countable (with CTLZ trip count).
1258 /// If CTLZ inserted as a new trip count returns true; otherwise, returns false.
1259 bool LoopIdiomRecognize::recognizeAndInsertCTLZ() {
1260 // Give up if the loop has multiple blocks or multiple backedges.
1261 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1264 Instruction *CntInst, *DefX;
1265 PHINode *CntPhi, *PhiX;
1266 if (!detectCTLZIdiom(CurLoop, PhiX, CntInst, CntPhi, DefX))
1269 bool IsCntPhiUsedOutsideLoop = false;
1270 for (User *U : CntPhi->users())
1271 if (!CurLoop->contains(dyn_cast<Instruction>(U))) {
1272 IsCntPhiUsedOutsideLoop = true;
1275 bool IsCntInstUsedOutsideLoop = false;
1276 for (User *U : CntInst->users())
1277 if (!CurLoop->contains(dyn_cast<Instruction>(U))) {
1278 IsCntInstUsedOutsideLoop = true;
1281 // If both CntInst and CntPhi are used outside the loop the profitability
1283 if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop)
1286 // For some CPUs result of CTLZ(X) intrinsic is undefined
1287 // when X is 0. If we can not guarantee X != 0, we need to check this
1289 bool ZeroCheck = false;
1290 // It is safe to assume Preheader exist as it was checked in
1291 // parent function RunOnLoop.
1292 BasicBlock *PH = CurLoop->getLoopPreheader();
1293 Value *InitX = PhiX->getIncomingValueForBlock(PH);
1294 // If we check X != 0 before entering the loop we don't need a zero
1295 // check in CTLZ intrinsic, but only if Cnt Phi is not used outside of the
1296 // loop (if it is used we count CTLZ(X >> 1)).
1297 if (!IsCntPhiUsedOutsideLoop)
1298 if (BasicBlock *PreCondBB = PH->getSinglePredecessor())
1299 if (BranchInst *PreCondBr =
1300 dyn_cast<BranchInst>(PreCondBB->getTerminator())) {
1301 if (matchCondition(PreCondBr, PH) == InitX)
1305 // Check if CTLZ intrinsic is profitable. Assume it is always profitable
1306 // if we delete the loop (the loop has only 6 instructions):
1307 // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ]
1308 // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ]
1309 // %shr = ashr %n.addr.0, 1
1310 // %tobool = icmp eq %shr, 0
1311 // %inc = add nsw %i.0, 1
1314 IRBuilder<> Builder(PH->getTerminator());
1315 SmallVector<const Value *, 2> Ops =
1316 {InitX, ZeroCheck ? Builder.getTrue() : Builder.getFalse()};
1317 ArrayRef<const Value *> Args(Ops);
1318 if (CurLoop->getHeader()->size() != 6 &&
1319 TTI->getIntrinsicCost(Intrinsic::ctlz, InitX->getType(), Args) >
1320 TargetTransformInfo::TCC_Basic)
1323 const DebugLoc DL = DefX->getDebugLoc();
1324 transformLoopToCountable(PH, CntInst, CntPhi, InitX, DL, ZeroCheck,
1325 IsCntPhiUsedOutsideLoop);
1329 /// Recognizes a population count idiom in a non-countable loop.
1331 /// If detected, transforms the relevant code to issue the popcount intrinsic
1332 /// function call, and returns true; otherwise, returns false.
1333 bool LoopIdiomRecognize::recognizePopcount() {
1334 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
1337 // Counting population are usually conducted by few arithmetic instructions.
1338 // Such instructions can be easily "absorbed" by vacant slots in a
1339 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
1340 // in a compact loop.
1342 // Give up if the loop has multiple blocks or multiple backedges.
1343 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1346 BasicBlock *LoopBody = *(CurLoop->block_begin());
1347 if (LoopBody->size() >= 20) {
1348 // The loop is too big, bail out.
1352 // It should have a preheader containing nothing but an unconditional branch.
1353 BasicBlock *PH = CurLoop->getLoopPreheader();
1354 if (!PH || &PH->front() != PH->getTerminator())
1356 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
1357 if (!EntryBI || EntryBI->isConditional())
1360 // It should have a precondition block where the generated popcount instrinsic
1361 // function can be inserted.
1362 auto *PreCondBB = PH->getSinglePredecessor();
1365 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1366 if (!PreCondBI || PreCondBI->isUnconditional())
1369 Instruction *CntInst;
1372 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
1375 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
1379 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
1380 const DebugLoc &DL) {
1381 Value *Ops[] = {Val};
1382 Type *Tys[] = {Val->getType()};
1384 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
1385 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
1386 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
1387 CI->setDebugLoc(DL);
1392 static CallInst *createCTLZIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
1393 const DebugLoc &DL, bool ZeroCheck) {
1394 Value *Ops[] = {Val, ZeroCheck ? IRBuilder.getTrue() : IRBuilder.getFalse()};
1395 Type *Tys[] = {Val->getType()};
1397 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
1398 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctlz, Tys);
1399 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
1400 CI->setDebugLoc(DL);
1405 /// Transform the following loop:
1407 /// CntPhi = PHI [Cnt0, CntInst]
1408 /// PhiX = PHI [InitX, DefX]
1409 /// CntInst = CntPhi + 1
1410 /// DefX = PhiX >> 1
1412 /// Br: loop if (DefX != 0)
1413 /// Use(CntPhi) or Use(CntInst)
1416 /// If CntPhi used outside the loop:
1417 /// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1)
1418 /// Count = CountPrev + 1
1420 /// Count = BitWidth(InitX) - CTLZ(InitX)
1422 /// CntPhi = PHI [Cnt0, CntInst]
1423 /// PhiX = PHI [InitX, DefX]
1424 /// PhiCount = PHI [Count, Dec]
1425 /// CntInst = CntPhi + 1
1426 /// DefX = PhiX >> 1
1427 /// Dec = PhiCount - 1
1429 /// Br: loop if (Dec != 0)
1430 /// Use(CountPrev + Cnt0) // Use(CntPhi)
1432 /// Use(Count + Cnt0) // Use(CntInst)
1434 /// If LOOP_BODY is empty the loop will be deleted.
1435 /// If CntInst and DefX are not used in LOOP_BODY they will be removed.
1436 void LoopIdiomRecognize::transformLoopToCountable(
1437 BasicBlock *Preheader, Instruction *CntInst, PHINode *CntPhi, Value *InitX,
1438 const DebugLoc DL, bool ZeroCheck, bool IsCntPhiUsedOutsideLoop) {
1439 BranchInst *PreheaderBr = dyn_cast<BranchInst>(Preheader->getTerminator());
1441 // Step 1: Insert the CTLZ instruction at the end of the preheader block
1442 // Count = BitWidth - CTLZ(InitX);
1443 // If there are uses of CntPhi create:
1444 // CountPrev = BitWidth - CTLZ(InitX >> 1);
1445 IRBuilder<> Builder(PreheaderBr);
1446 Builder.SetCurrentDebugLocation(DL);
1447 Value *CTLZ, *Count, *CountPrev, *NewCount, *InitXNext;
1449 if (IsCntPhiUsedOutsideLoop)
1450 InitXNext = Builder.CreateAShr(InitX,
1451 ConstantInt::get(InitX->getType(), 1));
1454 CTLZ = createCTLZIntrinsic(Builder, InitXNext, DL, ZeroCheck);
1455 Count = Builder.CreateSub(
1456 ConstantInt::get(CTLZ->getType(),
1457 CTLZ->getType()->getIntegerBitWidth()),
1459 if (IsCntPhiUsedOutsideLoop) {
1461 Count = Builder.CreateAdd(
1463 ConstantInt::get(CountPrev->getType(), 1));
1465 if (IsCntPhiUsedOutsideLoop)
1466 NewCount = Builder.CreateZExtOrTrunc(CountPrev,
1467 cast<IntegerType>(CntInst->getType()));
1469 NewCount = Builder.CreateZExtOrTrunc(Count,
1470 cast<IntegerType>(CntInst->getType()));
1472 // If the CTLZ counter's initial value is not zero, insert Add Inst.
1473 Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader);
1474 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
1475 if (!InitConst || !InitConst->isZero())
1476 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
1478 // Step 2: Insert new IV and loop condition:
1481 // PhiCount = PHI [Count, Dec]
1483 // Dec = PhiCount - 1
1485 // Br: loop if (Dec != 0)
1486 BasicBlock *Body = *(CurLoop->block_begin());
1487 auto *LbBr = dyn_cast<BranchInst>(Body->getTerminator());
1488 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
1489 Type *Ty = Count->getType();
1491 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
1493 Builder.SetInsertPoint(LbCond);
1494 Instruction *TcDec = cast<Instruction>(
1495 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
1496 "tcdec", false, true));
1498 TcPhi->addIncoming(Count, Preheader);
1499 TcPhi->addIncoming(TcDec, Body);
1501 CmpInst::Predicate Pred =
1502 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ;
1503 LbCond->setPredicate(Pred);
1504 LbCond->setOperand(0, TcDec);
1505 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
1507 // Step 3: All the references to the original counter outside
1508 // the loop are replaced with the NewCount -- the value returned from
1509 // __builtin_ctlz(x).
1510 if (IsCntPhiUsedOutsideLoop)
1511 CntPhi->replaceUsesOutsideBlock(NewCount, Body);
1513 CntInst->replaceUsesOutsideBlock(NewCount, Body);
1515 // step 4: Forget the "non-computable" trip-count SCEV associated with the
1516 // loop. The loop would otherwise not be deleted even if it becomes empty.
1517 SE->forgetLoop(CurLoop);
1520 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
1521 Instruction *CntInst,
1522 PHINode *CntPhi, Value *Var) {
1523 BasicBlock *PreHead = CurLoop->getLoopPreheader();
1524 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1525 const DebugLoc DL = CntInst->getDebugLoc();
1527 // Assuming before transformation, the loop is following:
1528 // if (x) // the precondition
1529 // do { cnt++; x &= x - 1; } while(x);
1531 // Step 1: Insert the ctpop instruction at the end of the precondition block
1532 IRBuilder<> Builder(PreCondBr);
1533 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
1535 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
1536 NewCount = PopCntZext =
1537 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
1539 if (NewCount != PopCnt)
1540 (cast<Instruction>(NewCount))->setDebugLoc(DL);
1542 // TripCnt is exactly the number of iterations the loop has
1545 // If the population counter's initial value is not zero, insert Add Inst.
1546 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
1547 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
1548 if (!InitConst || !InitConst->isZero()) {
1549 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
1550 (cast<Instruction>(NewCount))->setDebugLoc(DL);
1554 // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
1555 // "if (NewCount == 0) loop-exit". Without this change, the intrinsic
1556 // function would be partial dead code, and downstream passes will drag
1557 // it back from the precondition block to the preheader.
1559 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
1561 Value *Opnd0 = PopCntZext;
1562 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
1563 if (PreCond->getOperand(0) != Var)
1564 std::swap(Opnd0, Opnd1);
1566 ICmpInst *NewPreCond = cast<ICmpInst>(
1567 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
1568 PreCondBr->setCondition(NewPreCond);
1570 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
1573 // Step 3: Note that the population count is exactly the trip count of the
1574 // loop in question, which enable us to to convert the loop from noncountable
1575 // loop into a countable one. The benefit is twofold:
1577 // - If the loop only counts population, the entire loop becomes dead after
1578 // the transformation. It is a lot easier to prove a countable loop dead
1579 // than to prove a noncountable one. (In some C dialects, an infinite loop
1580 // isn't dead even if it computes nothing useful. In general, DCE needs
1581 // to prove a noncountable loop finite before safely delete it.)
1583 // - If the loop also performs something else, it remains alive.
1584 // Since it is transformed to countable form, it can be aggressively
1585 // optimized by some optimizations which are in general not applicable
1586 // to a noncountable loop.
1588 // After this step, this loop (conceptually) would look like following:
1589 // newcnt = __builtin_ctpop(x);
1592 // do { cnt++; x &= x-1; t--) } while (t > 0);
1593 BasicBlock *Body = *(CurLoop->block_begin());
1595 auto *LbBr = dyn_cast<BranchInst>(Body->getTerminator());
1596 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
1597 Type *Ty = TripCnt->getType();
1599 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
1601 Builder.SetInsertPoint(LbCond);
1602 Instruction *TcDec = cast<Instruction>(
1603 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
1604 "tcdec", false, true));
1606 TcPhi->addIncoming(TripCnt, PreHead);
1607 TcPhi->addIncoming(TcDec, Body);
1609 CmpInst::Predicate Pred =
1610 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
1611 LbCond->setPredicate(Pred);
1612 LbCond->setOperand(0, TcDec);
1613 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
1616 // Step 4: All the references to the original population counter outside
1617 // the loop are replaced with the NewCount -- the value returned from
1618 // __builtin_ctpop().
1619 CntInst->replaceUsesOutsideBlock(NewCount, Body);
1621 // step 5: Forget the "non-computable" trip-count SCEV associated with the
1622 // loop. The loop would otherwise not be deleted even if it becomes empty.
1623 SE->forgetLoop(CurLoop);