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/ADT/APInt.h"
41 #include "llvm/ADT/ArrayRef.h"
42 #include "llvm/ADT/DenseMap.h"
43 #include "llvm/ADT/MapVector.h"
44 #include "llvm/ADT/SetVector.h"
45 #include "llvm/ADT/SmallPtrSet.h"
46 #include "llvm/ADT/SmallVector.h"
47 #include "llvm/ADT/Statistic.h"
48 #include "llvm/ADT/StringRef.h"
49 #include "llvm/Analysis/AliasAnalysis.h"
50 #include "llvm/Analysis/LoopAccessAnalysis.h"
51 #include "llvm/Analysis/LoopInfo.h"
52 #include "llvm/Analysis/LoopPass.h"
53 #include "llvm/Analysis/MemoryLocation.h"
54 #include "llvm/Analysis/ScalarEvolution.h"
55 #include "llvm/Analysis/ScalarEvolutionExpander.h"
56 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
57 #include "llvm/Analysis/TargetLibraryInfo.h"
58 #include "llvm/Analysis/TargetTransformInfo.h"
59 #include "llvm/Transforms/Utils/Local.h"
60 #include "llvm/Analysis/ValueTracking.h"
61 #include "llvm/IR/Attributes.h"
62 #include "llvm/IR/BasicBlock.h"
63 #include "llvm/IR/Constant.h"
64 #include "llvm/IR/Constants.h"
65 #include "llvm/IR/DataLayout.h"
66 #include "llvm/IR/DebugLoc.h"
67 #include "llvm/IR/DerivedTypes.h"
68 #include "llvm/IR/Dominators.h"
69 #include "llvm/IR/GlobalValue.h"
70 #include "llvm/IR/GlobalVariable.h"
71 #include "llvm/IR/IRBuilder.h"
72 #include "llvm/IR/InstrTypes.h"
73 #include "llvm/IR/Instruction.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/IntrinsicInst.h"
76 #include "llvm/IR/Intrinsics.h"
77 #include "llvm/IR/LLVMContext.h"
78 #include "llvm/IR/Module.h"
79 #include "llvm/IR/PassManager.h"
80 #include "llvm/IR/Type.h"
81 #include "llvm/IR/User.h"
82 #include "llvm/IR/Value.h"
83 #include "llvm/IR/ValueHandle.h"
84 #include "llvm/Pass.h"
85 #include "llvm/Support/Casting.h"
86 #include "llvm/Support/CommandLine.h"
87 #include "llvm/Support/Debug.h"
88 #include "llvm/Support/raw_ostream.h"
89 #include "llvm/Transforms/Scalar.h"
90 #include "llvm/Transforms/Scalar/LoopIdiomRecognize.h"
91 #include "llvm/Transforms/Utils/BuildLibCalls.h"
92 #include "llvm/Transforms/Utils/LoopUtils.h"
101 #define DEBUG_TYPE "loop-idiom"
103 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
104 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
106 static cl::opt<bool> UseLIRCodeSizeHeurs(
107 "use-lir-code-size-heurs",
108 cl::desc("Use loop idiom recognition code size heuristics when compiling"
110 cl::init(true), cl::Hidden);
114 class LoopIdiomRecognize {
115 Loop *CurLoop = nullptr;
120 TargetLibraryInfo *TLI;
121 const TargetTransformInfo *TTI;
122 const DataLayout *DL;
123 bool ApplyCodeSizeHeuristics;
126 explicit LoopIdiomRecognize(AliasAnalysis *AA, DominatorTree *DT,
127 LoopInfo *LI, ScalarEvolution *SE,
128 TargetLibraryInfo *TLI,
129 const TargetTransformInfo *TTI,
130 const DataLayout *DL)
131 : AA(AA), DT(DT), LI(LI), SE(SE), TLI(TLI), TTI(TTI), DL(DL) {}
133 bool runOnLoop(Loop *L);
136 using StoreList = SmallVector<StoreInst *, 8>;
137 using StoreListMap = MapVector<Value *, StoreList>;
139 StoreListMap StoreRefsForMemset;
140 StoreListMap StoreRefsForMemsetPattern;
141 StoreList StoreRefsForMemcpy;
143 bool HasMemsetPattern;
146 /// Return code for isLegalStore()
147 enum LegalStoreKind {
152 UnorderedAtomicMemcpy,
153 DontUse // Dummy retval never to be used. Allows catching errors in retval
157 /// \name Countable Loop Idiom Handling
160 bool runOnCountableLoop();
161 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
162 SmallVectorImpl<BasicBlock *> &ExitBlocks);
164 void collectStores(BasicBlock *BB);
165 LegalStoreKind isLegalStore(StoreInst *SI);
166 bool processLoopStores(SmallVectorImpl<StoreInst *> &SL, const SCEV *BECount,
168 bool processLoopMemSet(MemSetInst *MSI, const SCEV *BECount);
170 bool processLoopStridedStore(Value *DestPtr, unsigned StoreSize,
171 unsigned StoreAlignment, Value *StoredVal,
172 Instruction *TheStore,
173 SmallPtrSetImpl<Instruction *> &Stores,
174 const SCEVAddRecExpr *Ev, const SCEV *BECount,
175 bool NegStride, bool IsLoopMemset = false);
176 bool processLoopStoreOfLoopLoad(StoreInst *SI, const SCEV *BECount);
177 bool avoidLIRForMultiBlockLoop(bool IsMemset = false,
178 bool IsLoopMemset = false);
181 /// \name Noncountable Loop Idiom Handling
184 bool runOnNoncountableLoop();
186 bool recognizePopcount();
187 void transformLoopToPopcount(BasicBlock *PreCondBB, Instruction *CntInst,
188 PHINode *CntPhi, Value *Var);
189 bool recognizeAndInsertCTLZ();
190 void transformLoopToCountable(BasicBlock *PreCondBB, Instruction *CntInst,
191 PHINode *CntPhi, Value *Var, Instruction *DefX,
192 const DebugLoc &DL, bool ZeroCheck,
193 bool IsCntPhiUsedOutsideLoop);
198 class LoopIdiomRecognizeLegacyPass : public LoopPass {
202 explicit LoopIdiomRecognizeLegacyPass() : LoopPass(ID) {
203 initializeLoopIdiomRecognizeLegacyPassPass(
204 *PassRegistry::getPassRegistry());
207 bool runOnLoop(Loop *L, LPPassManager &LPM) override {
211 AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
212 DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
213 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
214 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
215 TargetLibraryInfo *TLI =
216 &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
217 const TargetTransformInfo *TTI =
218 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(
219 *L->getHeader()->getParent());
220 const DataLayout *DL = &L->getHeader()->getModule()->getDataLayout();
222 LoopIdiomRecognize LIR(AA, DT, LI, SE, TLI, TTI, DL);
223 return LIR.runOnLoop(L);
226 /// This transformation requires natural loop information & requires that
227 /// loop preheaders be inserted into the CFG.
228 void getAnalysisUsage(AnalysisUsage &AU) const override {
229 AU.addRequired<TargetLibraryInfoWrapperPass>();
230 AU.addRequired<TargetTransformInfoWrapperPass>();
231 getLoopAnalysisUsage(AU);
235 } // end anonymous namespace
237 char LoopIdiomRecognizeLegacyPass::ID = 0;
239 PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM,
240 LoopStandardAnalysisResults &AR,
242 const auto *DL = &L.getHeader()->getModule()->getDataLayout();
244 LoopIdiomRecognize LIR(&AR.AA, &AR.DT, &AR.LI, &AR.SE, &AR.TLI, &AR.TTI, DL);
245 if (!LIR.runOnLoop(&L))
246 return PreservedAnalyses::all();
248 return getLoopPassPreservedAnalyses();
251 INITIALIZE_PASS_BEGIN(LoopIdiomRecognizeLegacyPass, "loop-idiom",
252 "Recognize loop idioms", false, false)
253 INITIALIZE_PASS_DEPENDENCY(LoopPass)
254 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
255 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
256 INITIALIZE_PASS_END(LoopIdiomRecognizeLegacyPass, "loop-idiom",
257 "Recognize loop idioms", false, false)
259 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognizeLegacyPass(); }
261 static void deleteDeadInstruction(Instruction *I) {
262 I->replaceAllUsesWith(UndefValue::get(I->getType()));
263 I->eraseFromParent();
266 //===----------------------------------------------------------------------===//
268 // Implementation of LoopIdiomRecognize
270 //===----------------------------------------------------------------------===//
272 bool LoopIdiomRecognize::runOnLoop(Loop *L) {
274 // If the loop could not be converted to canonical form, it must have an
275 // indirectbr in it, just give up.
276 if (!L->getLoopPreheader())
279 // Disable loop idiom recognition if the function's name is a common idiom.
280 StringRef Name = L->getHeader()->getParent()->getName();
281 if (Name == "memset" || Name == "memcpy")
284 // Determine if code size heuristics need to be applied.
285 ApplyCodeSizeHeuristics =
286 L->getHeader()->getParent()->optForSize() && UseLIRCodeSizeHeurs;
288 HasMemset = TLI->has(LibFunc_memset);
289 HasMemsetPattern = TLI->has(LibFunc_memset_pattern16);
290 HasMemcpy = TLI->has(LibFunc_memcpy);
292 if (HasMemset || HasMemsetPattern || HasMemcpy)
293 if (SE->hasLoopInvariantBackedgeTakenCount(L))
294 return runOnCountableLoop();
296 return runOnNoncountableLoop();
299 bool LoopIdiomRecognize::runOnCountableLoop() {
300 const SCEV *BECount = SE->getBackedgeTakenCount(CurLoop);
301 assert(!isa<SCEVCouldNotCompute>(BECount) &&
302 "runOnCountableLoop() called on a loop without a predictable"
303 "backedge-taken count");
305 // If this loop executes exactly one time, then it should be peeled, not
306 // optimized by this pass.
307 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
308 if (BECst->getAPInt() == 0)
311 SmallVector<BasicBlock *, 8> ExitBlocks;
312 CurLoop->getUniqueExitBlocks(ExitBlocks);
314 LLVM_DEBUG(dbgs() << "loop-idiom Scanning: F["
315 << CurLoop->getHeader()->getParent()->getName()
316 << "] Loop %" << CurLoop->getHeader()->getName() << "\n");
318 bool MadeChange = false;
320 // The following transforms hoist stores/memsets into the loop pre-header.
321 // Give up if the loop has instructions may throw.
322 LoopSafetyInfo SafetyInfo;
323 computeLoopSafetyInfo(&SafetyInfo, CurLoop);
324 if (SafetyInfo.MayThrow)
327 // Scan all the blocks in the loop that are not in subloops.
328 for (auto *BB : CurLoop->getBlocks()) {
329 // Ignore blocks in subloops.
330 if (LI->getLoopFor(BB) != CurLoop)
333 MadeChange |= runOnLoopBlock(BB, BECount, ExitBlocks);
338 static APInt getStoreStride(const SCEVAddRecExpr *StoreEv) {
339 const SCEVConstant *ConstStride = cast<SCEVConstant>(StoreEv->getOperand(1));
340 return ConstStride->getAPInt();
343 /// getMemSetPatternValue - If a strided store of the specified value is safe to
344 /// turn into a memset_pattern16, return a ConstantArray of 16 bytes that should
345 /// be passed in. Otherwise, return null.
347 /// Note that we don't ever attempt to use memset_pattern8 or 4, because these
348 /// just replicate their input array and then pass on to memset_pattern16.
349 static Constant *getMemSetPatternValue(Value *V, const DataLayout *DL) {
350 // If the value isn't a constant, we can't promote it to being in a constant
351 // array. We could theoretically do a store to an alloca or something, but
352 // that doesn't seem worthwhile.
353 Constant *C = dyn_cast<Constant>(V);
357 // Only handle simple values that are a power of two bytes in size.
358 uint64_t Size = DL->getTypeSizeInBits(V->getType());
359 if (Size == 0 || (Size & 7) || (Size & (Size - 1)))
362 // Don't care enough about darwin/ppc to implement this.
363 if (DL->isBigEndian())
366 // Convert to size in bytes.
369 // TODO: If CI is larger than 16-bytes, we can try slicing it in half to see
370 // if the top and bottom are the same (e.g. for vectors and large integers).
374 // If the constant is exactly 16 bytes, just use it.
378 // Otherwise, we'll use an array of the constants.
379 unsigned ArraySize = 16 / Size;
380 ArrayType *AT = ArrayType::get(V->getType(), ArraySize);
381 return ConstantArray::get(AT, std::vector<Constant *>(ArraySize, C));
384 LoopIdiomRecognize::LegalStoreKind
385 LoopIdiomRecognize::isLegalStore(StoreInst *SI) {
386 // Don't touch volatile stores.
387 if (SI->isVolatile())
388 return LegalStoreKind::None;
389 // We only want simple or unordered-atomic stores.
390 if (!SI->isUnordered())
391 return LegalStoreKind::None;
393 // Don't convert stores of non-integral pointer types to memsets (which stores
395 if (DL->isNonIntegralPointerType(SI->getValueOperand()->getType()))
396 return LegalStoreKind::None;
398 // Avoid merging nontemporal stores.
399 if (SI->getMetadata(LLVMContext::MD_nontemporal))
400 return LegalStoreKind::None;
402 Value *StoredVal = SI->getValueOperand();
403 Value *StorePtr = SI->getPointerOperand();
405 // Reject stores that are so large that they overflow an unsigned.
406 uint64_t SizeInBits = DL->getTypeSizeInBits(StoredVal->getType());
407 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
408 return LegalStoreKind::None;
410 // See if the pointer expression is an AddRec like {base,+,1} on the current
411 // loop, which indicates a strided store. If we have something else, it's a
412 // random store we can't handle.
413 const SCEVAddRecExpr *StoreEv =
414 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
415 if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
416 return LegalStoreKind::None;
418 // Check to see if we have a constant stride.
419 if (!isa<SCEVConstant>(StoreEv->getOperand(1)))
420 return LegalStoreKind::None;
422 // See if the store can be turned into a memset.
424 // If the stored value is a byte-wise value (like i32 -1), then it may be
425 // turned into a memset of i8 -1, assuming that all the consecutive bytes
426 // are stored. A store of i32 0x01020304 can never be turned into a memset,
427 // but it can be turned into memset_pattern if the target supports it.
428 Value *SplatValue = isBytewiseValue(StoredVal);
429 Constant *PatternValue = nullptr;
431 // Note: memset and memset_pattern on unordered-atomic is yet not supported
432 bool UnorderedAtomic = SI->isUnordered() && !SI->isSimple();
434 // If we're allowed to form a memset, and the stored value would be
435 // acceptable for memset, use it.
436 if (!UnorderedAtomic && HasMemset && SplatValue &&
437 // Verify that the stored value is loop invariant. If not, we can't
438 // promote the memset.
439 CurLoop->isLoopInvariant(SplatValue)) {
440 // It looks like we can use SplatValue.
441 return LegalStoreKind::Memset;
442 } else if (!UnorderedAtomic && HasMemsetPattern &&
443 // Don't create memset_pattern16s with address spaces.
444 StorePtr->getType()->getPointerAddressSpace() == 0 &&
445 (PatternValue = getMemSetPatternValue(StoredVal, DL))) {
446 // It looks like we can use PatternValue!
447 return LegalStoreKind::MemsetPattern;
450 // Otherwise, see if the store can be turned into a memcpy.
452 // Check to see if the stride matches the size of the store. If so, then we
453 // know that every byte is touched in the loop.
454 APInt Stride = getStoreStride(StoreEv);
455 unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
456 if (StoreSize != Stride && StoreSize != -Stride)
457 return LegalStoreKind::None;
459 // The store must be feeding a non-volatile load.
460 LoadInst *LI = dyn_cast<LoadInst>(SI->getValueOperand());
462 // Only allow non-volatile loads
463 if (!LI || LI->isVolatile())
464 return LegalStoreKind::None;
465 // Only allow simple or unordered-atomic loads
466 if (!LI->isUnordered())
467 return LegalStoreKind::None;
469 // See if the pointer expression is an AddRec like {base,+,1} on the current
470 // loop, which indicates a strided load. If we have something else, it's a
471 // random load we can't handle.
472 const SCEVAddRecExpr *LoadEv =
473 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
474 if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
475 return LegalStoreKind::None;
477 // The store and load must share the same stride.
478 if (StoreEv->getOperand(1) != LoadEv->getOperand(1))
479 return LegalStoreKind::None;
481 // Success. This store can be converted into a memcpy.
482 UnorderedAtomic = UnorderedAtomic || LI->isAtomic();
483 return UnorderedAtomic ? LegalStoreKind::UnorderedAtomicMemcpy
484 : LegalStoreKind::Memcpy;
486 // This store can't be transformed into a memset/memcpy.
487 return LegalStoreKind::None;
490 void LoopIdiomRecognize::collectStores(BasicBlock *BB) {
491 StoreRefsForMemset.clear();
492 StoreRefsForMemsetPattern.clear();
493 StoreRefsForMemcpy.clear();
494 for (Instruction &I : *BB) {
495 StoreInst *SI = dyn_cast<StoreInst>(&I);
499 // Make sure this is a strided store with a constant stride.
500 switch (isLegalStore(SI)) {
501 case LegalStoreKind::None:
504 case LegalStoreKind::Memset: {
505 // Find the base pointer.
506 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), *DL);
507 StoreRefsForMemset[Ptr].push_back(SI);
509 case LegalStoreKind::MemsetPattern: {
510 // Find the base pointer.
511 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), *DL);
512 StoreRefsForMemsetPattern[Ptr].push_back(SI);
514 case LegalStoreKind::Memcpy:
515 case LegalStoreKind::UnorderedAtomicMemcpy:
516 StoreRefsForMemcpy.push_back(SI);
519 assert(false && "unhandled return value");
525 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
526 /// with the specified backedge count. This block is known to be in the current
527 /// loop and not in any subloops.
528 bool LoopIdiomRecognize::runOnLoopBlock(
529 BasicBlock *BB, const SCEV *BECount,
530 SmallVectorImpl<BasicBlock *> &ExitBlocks) {
531 // We can only promote stores in this block if they are unconditionally
532 // executed in the loop. For a block to be unconditionally executed, it has
533 // to dominate all the exit blocks of the loop. Verify this now.
534 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
535 if (!DT->dominates(BB, ExitBlocks[i]))
538 bool MadeChange = false;
539 // Look for store instructions, which may be optimized to memset/memcpy.
542 // Look for a single store or sets of stores with a common base, which can be
543 // optimized into a memset (memset_pattern). The latter most commonly happens
544 // with structs and handunrolled loops.
545 for (auto &SL : StoreRefsForMemset)
546 MadeChange |= processLoopStores(SL.second, BECount, true);
548 for (auto &SL : StoreRefsForMemsetPattern)
549 MadeChange |= processLoopStores(SL.second, BECount, false);
551 // Optimize the store into a memcpy, if it feeds an similarly strided load.
552 for (auto &SI : StoreRefsForMemcpy)
553 MadeChange |= processLoopStoreOfLoopLoad(SI, BECount);
555 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
556 Instruction *Inst = &*I++;
557 // Look for memset instructions, which may be optimized to a larger memset.
558 if (MemSetInst *MSI = dyn_cast<MemSetInst>(Inst)) {
559 WeakTrackingVH InstPtr(&*I);
560 if (!processLoopMemSet(MSI, BECount))
564 // If processing the memset invalidated our iterator, start over from the
575 /// processLoopStores - See if this store(s) can be promoted to a memset.
576 bool LoopIdiomRecognize::processLoopStores(SmallVectorImpl<StoreInst *> &SL,
579 // Try to find consecutive stores that can be transformed into memsets.
580 SetVector<StoreInst *> Heads, Tails;
581 SmallDenseMap<StoreInst *, StoreInst *> ConsecutiveChain;
583 // Do a quadratic search on all of the given stores and find
584 // all of the pairs of stores that follow each other.
585 SmallVector<unsigned, 16> IndexQueue;
586 for (unsigned i = 0, e = SL.size(); i < e; ++i) {
587 assert(SL[i]->isSimple() && "Expected only non-volatile stores.");
589 Value *FirstStoredVal = SL[i]->getValueOperand();
590 Value *FirstStorePtr = SL[i]->getPointerOperand();
591 const SCEVAddRecExpr *FirstStoreEv =
592 cast<SCEVAddRecExpr>(SE->getSCEV(FirstStorePtr));
593 APInt FirstStride = getStoreStride(FirstStoreEv);
594 unsigned FirstStoreSize = DL->getTypeStoreSize(SL[i]->getValueOperand()->getType());
596 // See if we can optimize just this store in isolation.
597 if (FirstStride == FirstStoreSize || -FirstStride == FirstStoreSize) {
602 Value *FirstSplatValue = nullptr;
603 Constant *FirstPatternValue = nullptr;
606 FirstSplatValue = isBytewiseValue(FirstStoredVal);
608 FirstPatternValue = getMemSetPatternValue(FirstStoredVal, DL);
610 assert((FirstSplatValue || FirstPatternValue) &&
611 "Expected either splat value or pattern value.");
614 // If a store has multiple consecutive store candidates, search Stores
615 // array according to the sequence: from i+1 to e, then from i-1 to 0.
616 // This is because usually pairing with immediate succeeding or preceding
617 // candidate create the best chance to find memset opportunity.
619 for (j = i + 1; j < e; ++j)
620 IndexQueue.push_back(j);
621 for (j = i; j > 0; --j)
622 IndexQueue.push_back(j - 1);
624 for (auto &k : IndexQueue) {
625 assert(SL[k]->isSimple() && "Expected only non-volatile stores.");
626 Value *SecondStorePtr = SL[k]->getPointerOperand();
627 const SCEVAddRecExpr *SecondStoreEv =
628 cast<SCEVAddRecExpr>(SE->getSCEV(SecondStorePtr));
629 APInt SecondStride = getStoreStride(SecondStoreEv);
631 if (FirstStride != SecondStride)
634 Value *SecondStoredVal = SL[k]->getValueOperand();
635 Value *SecondSplatValue = nullptr;
636 Constant *SecondPatternValue = nullptr;
639 SecondSplatValue = isBytewiseValue(SecondStoredVal);
641 SecondPatternValue = getMemSetPatternValue(SecondStoredVal, DL);
643 assert((SecondSplatValue || SecondPatternValue) &&
644 "Expected either splat value or pattern value.");
646 if (isConsecutiveAccess(SL[i], SL[k], *DL, *SE, false)) {
648 if (FirstSplatValue != SecondSplatValue)
651 if (FirstPatternValue != SecondPatternValue)
656 ConsecutiveChain[SL[i]] = SL[k];
662 // We may run into multiple chains that merge into a single chain. We mark the
663 // stores that we transformed so that we don't visit the same store twice.
664 SmallPtrSet<Value *, 16> TransformedStores;
665 bool Changed = false;
667 // For stores that start but don't end a link in the chain:
668 for (SetVector<StoreInst *>::iterator it = Heads.begin(), e = Heads.end();
670 if (Tails.count(*it))
673 // We found a store instr that starts a chain. Now follow the chain and try
675 SmallPtrSet<Instruction *, 8> AdjacentStores;
678 StoreInst *HeadStore = I;
679 unsigned StoreSize = 0;
681 // Collect the chain into a list.
682 while (Tails.count(I) || Heads.count(I)) {
683 if (TransformedStores.count(I))
685 AdjacentStores.insert(I);
687 StoreSize += DL->getTypeStoreSize(I->getValueOperand()->getType());
688 // Move to the next value in the chain.
689 I = ConsecutiveChain[I];
692 Value *StoredVal = HeadStore->getValueOperand();
693 Value *StorePtr = HeadStore->getPointerOperand();
694 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
695 APInt Stride = getStoreStride(StoreEv);
697 // Check to see if the stride matches the size of the stores. If so, then
698 // we know that every byte is touched in the loop.
699 if (StoreSize != Stride && StoreSize != -Stride)
702 bool NegStride = StoreSize == -Stride;
704 if (processLoopStridedStore(StorePtr, StoreSize, HeadStore->getAlignment(),
705 StoredVal, HeadStore, AdjacentStores, StoreEv,
706 BECount, NegStride)) {
707 TransformedStores.insert(AdjacentStores.begin(), AdjacentStores.end());
715 /// processLoopMemSet - See if this memset can be promoted to a large memset.
716 bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
717 const SCEV *BECount) {
718 // We can only handle non-volatile memsets with a constant size.
719 if (MSI->isVolatile() || !isa<ConstantInt>(MSI->getLength()))
722 // If we're not allowed to hack on memset, we fail.
726 Value *Pointer = MSI->getDest();
728 // See if the pointer expression is an AddRec like {base,+,1} on the current
729 // loop, which indicates a strided store. If we have something else, it's a
730 // random store we can't handle.
731 const SCEVAddRecExpr *Ev = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Pointer));
732 if (!Ev || Ev->getLoop() != CurLoop || !Ev->isAffine())
735 // Reject memsets that are so large that they overflow an unsigned.
736 uint64_t SizeInBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
737 if ((SizeInBytes >> 32) != 0)
740 // Check to see if the stride matches the size of the memset. If so, then we
741 // know that every byte is touched in the loop.
742 const SCEVConstant *ConstStride = dyn_cast<SCEVConstant>(Ev->getOperand(1));
746 APInt Stride = ConstStride->getAPInt();
747 if (SizeInBytes != Stride && SizeInBytes != -Stride)
750 // Verify that the memset value is loop invariant. If not, we can't promote
752 Value *SplatValue = MSI->getValue();
753 if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue))
756 SmallPtrSet<Instruction *, 1> MSIs;
758 bool NegStride = SizeInBytes == -Stride;
759 return processLoopStridedStore(Pointer, (unsigned)SizeInBytes,
760 MSI->getDestAlignment(), SplatValue, MSI, MSIs,
761 Ev, BECount, NegStride, /*IsLoopMemset=*/true);
764 /// mayLoopAccessLocation - Return true if the specified loop might access the
765 /// specified pointer location, which is a loop-strided access. The 'Access'
766 /// argument specifies what the verboten forms of access are (read or write).
768 mayLoopAccessLocation(Value *Ptr, ModRefInfo Access, Loop *L,
769 const SCEV *BECount, unsigned StoreSize,
771 SmallPtrSetImpl<Instruction *> &IgnoredStores) {
772 // Get the location that may be stored across the loop. Since the access is
773 // strided positively through memory, we say that the modified location starts
774 // at the pointer and has infinite size.
775 uint64_t AccessSize = MemoryLocation::UnknownSize;
777 // If the loop iterates a fixed number of times, we can refine the access size
778 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
779 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
780 AccessSize = (BECst->getValue()->getZExtValue() + 1) * StoreSize;
782 // TODO: For this to be really effective, we have to dive into the pointer
783 // operand in the store. Store to &A[i] of 100 will always return may alias
784 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
785 // which will then no-alias a store to &A[100].
786 MemoryLocation StoreLoc(Ptr, AccessSize);
788 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
790 for (Instruction &I : **BI)
791 if (IgnoredStores.count(&I) == 0 &&
793 intersectModRef(AA.getModRefInfo(&I, StoreLoc), Access)))
799 // If we have a negative stride, Start refers to the end of the memory location
800 // we're trying to memset. Therefore, we need to recompute the base pointer,
801 // which is just Start - BECount*Size.
802 static const SCEV *getStartForNegStride(const SCEV *Start, const SCEV *BECount,
803 Type *IntPtr, unsigned StoreSize,
804 ScalarEvolution *SE) {
805 const SCEV *Index = SE->getTruncateOrZeroExtend(BECount, IntPtr);
807 Index = SE->getMulExpr(Index, SE->getConstant(IntPtr, StoreSize),
809 return SE->getMinusSCEV(Start, Index);
812 /// Compute the number of bytes as a SCEV from the backedge taken count.
814 /// This also maps the SCEV into the provided type and tries to handle the
815 /// computation in a way that will fold cleanly.
816 static const SCEV *getNumBytes(const SCEV *BECount, Type *IntPtr,
817 unsigned StoreSize, Loop *CurLoop,
818 const DataLayout *DL, ScalarEvolution *SE) {
819 const SCEV *NumBytesS;
820 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
821 // pointer size if it isn't already.
823 // If we're going to need to zero extend the BE count, check if we can add
824 // one to it prior to zero extending without overflow. Provided this is safe,
825 // it allows better simplification of the +1.
826 if (DL->getTypeSizeInBits(BECount->getType()) <
827 DL->getTypeSizeInBits(IntPtr) &&
828 SE->isLoopEntryGuardedByCond(
829 CurLoop, ICmpInst::ICMP_NE, BECount,
830 SE->getNegativeSCEV(SE->getOne(BECount->getType())))) {
831 NumBytesS = SE->getZeroExtendExpr(
832 SE->getAddExpr(BECount, SE->getOne(BECount->getType()), SCEV::FlagNUW),
835 NumBytesS = SE->getAddExpr(SE->getTruncateOrZeroExtend(BECount, IntPtr),
836 SE->getOne(IntPtr), SCEV::FlagNUW);
839 // And scale it based on the store size.
840 if (StoreSize != 1) {
841 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
847 /// processLoopStridedStore - We see a strided store of some value. If we can
848 /// transform this into a memset or memset_pattern in the loop preheader, do so.
849 bool LoopIdiomRecognize::processLoopStridedStore(
850 Value *DestPtr, unsigned StoreSize, unsigned StoreAlignment,
851 Value *StoredVal, Instruction *TheStore,
852 SmallPtrSetImpl<Instruction *> &Stores, const SCEVAddRecExpr *Ev,
853 const SCEV *BECount, bool NegStride, bool IsLoopMemset) {
854 Value *SplatValue = isBytewiseValue(StoredVal);
855 Constant *PatternValue = nullptr;
858 PatternValue = getMemSetPatternValue(StoredVal, DL);
860 assert((SplatValue || PatternValue) &&
861 "Expected either splat value or pattern value.");
863 // The trip count of the loop and the base pointer of the addrec SCEV is
864 // guaranteed to be loop invariant, which means that it should dominate the
865 // header. This allows us to insert code for it in the preheader.
866 unsigned DestAS = DestPtr->getType()->getPointerAddressSpace();
867 BasicBlock *Preheader = CurLoop->getLoopPreheader();
868 IRBuilder<> Builder(Preheader->getTerminator());
869 SCEVExpander Expander(*SE, *DL, "loop-idiom");
871 Type *DestInt8PtrTy = Builder.getInt8PtrTy(DestAS);
872 Type *IntPtr = Builder.getIntPtrTy(*DL, DestAS);
874 const SCEV *Start = Ev->getStart();
875 // Handle negative strided loops.
877 Start = getStartForNegStride(Start, BECount, IntPtr, StoreSize, SE);
879 // TODO: ideally we should still be able to generate memset if SCEV expander
880 // is taught to generate the dependencies at the latest point.
881 if (!isSafeToExpand(Start, *SE))
884 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
885 // this into a memset in the loop preheader now if we want. However, this
886 // would be unsafe to do if there is anything else in the loop that may read
887 // or write to the aliased location. Check for any overlap by generating the
888 // base pointer and checking the region.
890 Expander.expandCodeFor(Start, DestInt8PtrTy, Preheader->getTerminator());
891 if (mayLoopAccessLocation(BasePtr, ModRefInfo::ModRef, CurLoop, BECount,
892 StoreSize, *AA, Stores)) {
894 // If we generated new code for the base pointer, clean up.
895 RecursivelyDeleteTriviallyDeadInstructions(BasePtr, TLI);
899 if (avoidLIRForMultiBlockLoop(/*IsMemset=*/true, IsLoopMemset))
902 // Okay, everything looks good, insert the memset.
904 const SCEV *NumBytesS =
905 getNumBytes(BECount, IntPtr, StoreSize, CurLoop, DL, SE);
907 // TODO: ideally we should still be able to generate memset if SCEV expander
908 // is taught to generate the dependencies at the latest point.
909 if (!isSafeToExpand(NumBytesS, *SE))
913 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
918 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, StoreAlignment);
920 // Everything is emitted in default address space
921 Type *Int8PtrTy = DestInt8PtrTy;
923 Module *M = TheStore->getModule();
925 M->getOrInsertFunction("memset_pattern16", Builder.getVoidTy(),
926 Int8PtrTy, Int8PtrTy, IntPtr);
927 inferLibFuncAttributes(*M->getFunction("memset_pattern16"), *TLI);
929 // Otherwise we should form a memset_pattern16. PatternValue is known to be
930 // an constant array of 16-bytes. Plop the value into a mergable global.
931 GlobalVariable *GV = new GlobalVariable(*M, PatternValue->getType(), true,
932 GlobalValue::PrivateLinkage,
933 PatternValue, ".memset_pattern");
934 GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); // Ok to merge these.
935 GV->setAlignment(16);
936 Value *PatternPtr = ConstantExpr::getBitCast(GV, Int8PtrTy);
937 NewCall = Builder.CreateCall(MSP, {BasePtr, PatternPtr, NumBytes});
940 LLVM_DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
941 << " from store to: " << *Ev << " at: " << *TheStore
943 NewCall->setDebugLoc(TheStore->getDebugLoc());
945 // Okay, the memset has been formed. Zap the original store and anything that
947 for (auto *I : Stores)
948 deleteDeadInstruction(I);
953 /// If the stored value is a strided load in the same loop with the same stride
954 /// this may be transformable into a memcpy. This kicks in for stuff like
955 /// for (i) A[i] = B[i];
956 bool LoopIdiomRecognize::processLoopStoreOfLoopLoad(StoreInst *SI,
957 const SCEV *BECount) {
958 assert(SI->isUnordered() && "Expected only non-volatile non-ordered stores.");
960 Value *StorePtr = SI->getPointerOperand();
961 const SCEVAddRecExpr *StoreEv = cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
962 APInt Stride = getStoreStride(StoreEv);
963 unsigned StoreSize = DL->getTypeStoreSize(SI->getValueOperand()->getType());
964 bool NegStride = StoreSize == -Stride;
966 // The store must be feeding a non-volatile load.
967 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
968 assert(LI->isUnordered() && "Expected only non-volatile non-ordered loads.");
970 // See if the pointer expression is an AddRec like {base,+,1} on the current
971 // loop, which indicates a strided load. If we have something else, it's a
972 // random load we can't handle.
973 const SCEVAddRecExpr *LoadEv =
974 cast<SCEVAddRecExpr>(SE->getSCEV(LI->getPointerOperand()));
976 // The trip count of the loop and the base pointer of the addrec SCEV is
977 // guaranteed to be loop invariant, which means that it should dominate the
978 // header. This allows us to insert code for it in the preheader.
979 BasicBlock *Preheader = CurLoop->getLoopPreheader();
980 IRBuilder<> Builder(Preheader->getTerminator());
981 SCEVExpander Expander(*SE, *DL, "loop-idiom");
983 const SCEV *StrStart = StoreEv->getStart();
984 unsigned StrAS = SI->getPointerAddressSpace();
985 Type *IntPtrTy = Builder.getIntPtrTy(*DL, StrAS);
987 // Handle negative strided loops.
989 StrStart = getStartForNegStride(StrStart, BECount, IntPtrTy, StoreSize, SE);
991 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
992 // this into a memcpy in the loop preheader now if we want. However, this
993 // would be unsafe to do if there is anything else in the loop that may read
994 // or write the memory region we're storing to. This includes the load that
995 // feeds the stores. Check for an alias by generating the base address and
996 // checking everything.
997 Value *StoreBasePtr = Expander.expandCodeFor(
998 StrStart, Builder.getInt8PtrTy(StrAS), Preheader->getTerminator());
1000 SmallPtrSet<Instruction *, 1> Stores;
1002 if (mayLoopAccessLocation(StoreBasePtr, ModRefInfo::ModRef, CurLoop, BECount,
1003 StoreSize, *AA, Stores)) {
1005 // If we generated new code for the base pointer, clean up.
1006 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
1010 const SCEV *LdStart = LoadEv->getStart();
1011 unsigned LdAS = LI->getPointerAddressSpace();
1013 // Handle negative strided loops.
1015 LdStart = getStartForNegStride(LdStart, BECount, IntPtrTy, StoreSize, SE);
1017 // For a memcpy, we have to make sure that the input array is not being
1018 // mutated by the loop.
1019 Value *LoadBasePtr = Expander.expandCodeFor(
1020 LdStart, Builder.getInt8PtrTy(LdAS), Preheader->getTerminator());
1022 if (mayLoopAccessLocation(LoadBasePtr, ModRefInfo::Mod, CurLoop, BECount,
1023 StoreSize, *AA, Stores)) {
1025 // If we generated new code for the base pointer, clean up.
1026 RecursivelyDeleteTriviallyDeadInstructions(LoadBasePtr, TLI);
1027 RecursivelyDeleteTriviallyDeadInstructions(StoreBasePtr, TLI);
1031 if (avoidLIRForMultiBlockLoop())
1034 // Okay, everything is safe, we can transform this!
1036 const SCEV *NumBytesS =
1037 getNumBytes(BECount, IntPtrTy, StoreSize, CurLoop, DL, SE);
1040 Expander.expandCodeFor(NumBytesS, IntPtrTy, Preheader->getTerminator());
1042 CallInst *NewCall = nullptr;
1043 // Check whether to generate an unordered atomic memcpy:
1044 // If the load or store are atomic, then they must necessarily be unordered
1045 // by previous checks.
1046 if (!SI->isAtomic() && !LI->isAtomic())
1047 NewCall = Builder.CreateMemCpy(StoreBasePtr, SI->getAlignment(),
1048 LoadBasePtr, LI->getAlignment(), NumBytes);
1050 // We cannot allow unaligned ops for unordered load/store, so reject
1051 // anything where the alignment isn't at least the element size.
1052 unsigned Align = std::min(SI->getAlignment(), LI->getAlignment());
1053 if (Align < StoreSize)
1056 // If the element.atomic memcpy is not lowered into explicit
1057 // loads/stores later, then it will be lowered into an element-size
1058 // specific lib call. If the lib call doesn't exist for our store size, then
1059 // we shouldn't generate the memcpy.
1060 if (StoreSize > TTI->getAtomicMemIntrinsicMaxElementSize())
1064 // Note that unordered atomic loads/stores are *required* by the spec to
1065 // have an alignment but non-atomic loads/stores may not.
1066 NewCall = Builder.CreateElementUnorderedAtomicMemCpy(
1067 StoreBasePtr, SI->getAlignment(), LoadBasePtr, LI->getAlignment(),
1068 NumBytes, StoreSize);
1070 NewCall->setDebugLoc(SI->getDebugLoc());
1072 LLVM_DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
1073 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
1074 << " from store ptr=" << *StoreEv << " at: " << *SI
1077 // Okay, the memcpy has been formed. Zap the original store and anything that
1079 deleteDeadInstruction(SI);
1084 // When compiling for codesize we avoid idiom recognition for a multi-block loop
1085 // unless it is a loop_memset idiom or a memset/memcpy idiom in a nested loop.
1087 bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset,
1088 bool IsLoopMemset) {
1089 if (ApplyCodeSizeHeuristics && CurLoop->getNumBlocks() > 1) {
1090 if (!CurLoop->getParentLoop() && (!IsMemset || !IsLoopMemset)) {
1091 LLVM_DEBUG(dbgs() << " " << CurLoop->getHeader()->getParent()->getName()
1092 << " : LIR " << (IsMemset ? "Memset" : "Memcpy")
1093 << " avoided: multi-block top-level loop\n");
1101 bool LoopIdiomRecognize::runOnNoncountableLoop() {
1102 return recognizePopcount() || recognizeAndInsertCTLZ();
1105 /// Check if the given conditional branch is based on the comparison between
1106 /// a variable and zero, and if the variable is non-zero, the control yields to
1107 /// the loop entry. If the branch matches the behavior, the variable involved
1108 /// in the comparison is returned. This function will be called to see if the
1109 /// precondition and postcondition of the loop are in desirable form.
1110 static Value *matchCondition(BranchInst *BI, BasicBlock *LoopEntry) {
1111 if (!BI || !BI->isConditional())
1114 ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
1118 ConstantInt *CmpZero = dyn_cast<ConstantInt>(Cond->getOperand(1));
1119 if (!CmpZero || !CmpZero->isZero())
1122 ICmpInst::Predicate Pred = Cond->getPredicate();
1123 if ((Pred == ICmpInst::ICMP_NE && BI->getSuccessor(0) == LoopEntry) ||
1124 (Pred == ICmpInst::ICMP_EQ && BI->getSuccessor(1) == LoopEntry))
1125 return Cond->getOperand(0);
1130 // Check if the recurrence variable `VarX` is in the right form to create
1131 // the idiom. Returns the value coerced to a PHINode if so.
1132 static PHINode *getRecurrenceVar(Value *VarX, Instruction *DefX,
1133 BasicBlock *LoopEntry) {
1134 auto *PhiX = dyn_cast<PHINode>(VarX);
1135 if (PhiX && PhiX->getParent() == LoopEntry &&
1136 (PhiX->getOperand(0) == DefX || PhiX->getOperand(1) == DefX))
1141 /// Return true iff the idiom is detected in the loop.
1144 /// 1) \p CntInst is set to the instruction counting the population bit.
1145 /// 2) \p CntPhi is set to the corresponding phi node.
1146 /// 3) \p Var is set to the value whose population bits are being counted.
1148 /// The core idiom we are trying to detect is:
1151 /// goto loop-exit // the precondition of the loop
1152 /// cnt0 = init-val;
1154 /// x1 = phi (x0, x2);
1155 /// cnt1 = phi(cnt0, cnt2);
1157 /// cnt2 = cnt1 + 1;
1159 /// x2 = x1 & (x1 - 1);
1161 /// } while(x != 0);
1165 static bool detectPopcountIdiom(Loop *CurLoop, BasicBlock *PreCondBB,
1166 Instruction *&CntInst, PHINode *&CntPhi,
1168 // step 1: Check to see if the look-back branch match this pattern:
1169 // "if (a!=0) goto loop-entry".
1170 BasicBlock *LoopEntry;
1171 Instruction *DefX2, *CountInst;
1172 Value *VarX1, *VarX0;
1173 PHINode *PhiX, *CountPhi;
1175 DefX2 = CountInst = nullptr;
1176 VarX1 = VarX0 = nullptr;
1177 PhiX = CountPhi = nullptr;
1178 LoopEntry = *(CurLoop->block_begin());
1180 // step 1: Check if the loop-back branch is in desirable form.
1182 if (Value *T = matchCondition(
1183 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1184 DefX2 = dyn_cast<Instruction>(T);
1189 // step 2: detect instructions corresponding to "x2 = x1 & (x1 - 1)"
1191 if (!DefX2 || DefX2->getOpcode() != Instruction::And)
1194 BinaryOperator *SubOneOp;
1196 if ((SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(0))))
1197 VarX1 = DefX2->getOperand(1);
1199 VarX1 = DefX2->getOperand(0);
1200 SubOneOp = dyn_cast<BinaryOperator>(DefX2->getOperand(1));
1202 if (!SubOneOp || SubOneOp->getOperand(0) != VarX1)
1205 ConstantInt *Dec = dyn_cast<ConstantInt>(SubOneOp->getOperand(1));
1207 !((SubOneOp->getOpcode() == Instruction::Sub && Dec->isOne()) ||
1208 (SubOneOp->getOpcode() == Instruction::Add &&
1209 Dec->isMinusOne()))) {
1214 // step 3: Check the recurrence of variable X
1215 PhiX = getRecurrenceVar(VarX1, DefX2, LoopEntry);
1219 // step 4: Find the instruction which count the population: cnt2 = cnt1 + 1
1221 CountInst = nullptr;
1222 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
1223 IterE = LoopEntry->end();
1224 Iter != IterE; Iter++) {
1225 Instruction *Inst = &*Iter;
1226 if (Inst->getOpcode() != Instruction::Add)
1229 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
1230 if (!Inc || !Inc->isOne())
1233 PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry);
1237 // Check if the result of the instruction is live of the loop.
1238 bool LiveOutLoop = false;
1239 for (User *U : Inst->users()) {
1240 if ((cast<Instruction>(U))->getParent() != LoopEntry) {
1257 // step 5: check if the precondition is in this form:
1258 // "if (x != 0) goto loop-head ; else goto somewhere-we-don't-care;"
1260 auto *PreCondBr = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1261 Value *T = matchCondition(PreCondBr, CurLoop->getLoopPreheader());
1262 if (T != PhiX->getOperand(0) && T != PhiX->getOperand(1))
1265 CntInst = CountInst;
1273 /// Return true if the idiom is detected in the loop.
1276 /// 1) \p CntInst is set to the instruction Counting Leading Zeros (CTLZ)
1277 /// or nullptr if there is no such.
1278 /// 2) \p CntPhi is set to the corresponding phi node
1279 /// or nullptr if there is no such.
1280 /// 3) \p Var is set to the value whose CTLZ could be used.
1281 /// 4) \p DefX is set to the instruction calculating Loop exit condition.
1283 /// The core idiom we are trying to detect is:
1286 /// goto loop-exit // the precondition of the loop
1287 /// cnt0 = init-val;
1289 /// x = phi (x0, x.next); //PhiX
1290 /// cnt = phi(cnt0, cnt.next);
1292 /// cnt.next = cnt + 1;
1294 /// x.next = x >> 1; // DefX
1296 /// } while(x.next != 0);
1300 static bool detectCTLZIdiom(Loop *CurLoop, PHINode *&PhiX,
1301 Instruction *&CntInst, PHINode *&CntPhi,
1302 Instruction *&DefX) {
1303 BasicBlock *LoopEntry;
1304 Value *VarX = nullptr;
1310 LoopEntry = *(CurLoop->block_begin());
1312 // step 1: Check if the loop-back branch is in desirable form.
1313 if (Value *T = matchCondition(
1314 dyn_cast<BranchInst>(LoopEntry->getTerminator()), LoopEntry))
1315 DefX = dyn_cast<Instruction>(T);
1319 // step 2: detect instructions corresponding to "x.next = x >> 1"
1320 if (!DefX || (DefX->getOpcode() != Instruction::AShr &&
1321 DefX->getOpcode() != Instruction::LShr))
1323 ConstantInt *Shft = dyn_cast<ConstantInt>(DefX->getOperand(1));
1324 if (!Shft || !Shft->isOne())
1326 VarX = DefX->getOperand(0);
1328 // step 3: Check the recurrence of variable X
1329 PhiX = getRecurrenceVar(VarX, DefX, LoopEntry);
1333 // step 4: Find the instruction which count the CTLZ: cnt.next = cnt + 1
1334 // TODO: We can skip the step. If loop trip count is known (CTLZ),
1335 // then all uses of "cnt.next" could be optimized to the trip count
1336 // plus "cnt0". Currently it is not optimized.
1337 // This step could be used to detect POPCNT instruction:
1338 // cnt.next = cnt + (x.next & 1)
1339 for (BasicBlock::iterator Iter = LoopEntry->getFirstNonPHI()->getIterator(),
1340 IterE = LoopEntry->end();
1341 Iter != IterE; Iter++) {
1342 Instruction *Inst = &*Iter;
1343 if (Inst->getOpcode() != Instruction::Add)
1346 ConstantInt *Inc = dyn_cast<ConstantInt>(Inst->getOperand(1));
1347 if (!Inc || !Inc->isOne())
1350 PHINode *Phi = getRecurrenceVar(Inst->getOperand(0), Inst, LoopEntry);
1364 /// Recognize CTLZ idiom in a non-countable loop and convert the loop
1365 /// to countable (with CTLZ trip count).
1366 /// If CTLZ inserted as a new trip count returns true; otherwise, returns false.
1367 bool LoopIdiomRecognize::recognizeAndInsertCTLZ() {
1368 // Give up if the loop has multiple blocks or multiple backedges.
1369 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1372 Instruction *CntInst, *DefX;
1373 PHINode *CntPhi, *PhiX;
1374 if (!detectCTLZIdiom(CurLoop, PhiX, CntInst, CntPhi, DefX))
1377 bool IsCntPhiUsedOutsideLoop = false;
1378 for (User *U : CntPhi->users())
1379 if (!CurLoop->contains(cast<Instruction>(U))) {
1380 IsCntPhiUsedOutsideLoop = true;
1383 bool IsCntInstUsedOutsideLoop = false;
1384 for (User *U : CntInst->users())
1385 if (!CurLoop->contains(cast<Instruction>(U))) {
1386 IsCntInstUsedOutsideLoop = true;
1389 // If both CntInst and CntPhi are used outside the loop the profitability
1391 if (IsCntInstUsedOutsideLoop && IsCntPhiUsedOutsideLoop)
1394 // For some CPUs result of CTLZ(X) intrinsic is undefined
1395 // when X is 0. If we can not guarantee X != 0, we need to check this
1397 bool ZeroCheck = false;
1398 // It is safe to assume Preheader exist as it was checked in
1399 // parent function RunOnLoop.
1400 BasicBlock *PH = CurLoop->getLoopPreheader();
1401 Value *InitX = PhiX->getIncomingValueForBlock(PH);
1403 // Make sure the initial value can't be negative otherwise the ashr in the
1404 // loop might never reach zero which would make the loop infinite.
1405 if (DefX->getOpcode() == Instruction::AShr && !isKnownNonNegative(InitX, *DL))
1408 // If we are using the count instruction outside the loop, make sure we
1409 // have a zero check as a precondition. Without the check the loop would run
1410 // one iteration for before any check of the input value. This means 0 and 1
1411 // would have identical behavior in the original loop and thus
1412 if (!IsCntPhiUsedOutsideLoop) {
1413 auto *PreCondBB = PH->getSinglePredecessor();
1416 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1419 if (matchCondition(PreCondBI, PH) != InitX)
1424 // Check if CTLZ intrinsic is profitable. Assume it is always profitable
1425 // if we delete the loop (the loop has only 6 instructions):
1426 // %n.addr.0 = phi [ %n, %entry ], [ %shr, %while.cond ]
1427 // %i.0 = phi [ %i0, %entry ], [ %inc, %while.cond ]
1428 // %shr = ashr %n.addr.0, 1
1429 // %tobool = icmp eq %shr, 0
1430 // %inc = add nsw %i.0, 1
1433 const Value *Args[] =
1434 {InitX, ZeroCheck ? ConstantInt::getTrue(InitX->getContext())
1435 : ConstantInt::getFalse(InitX->getContext())};
1436 if (CurLoop->getHeader()->size() != 6 &&
1437 TTI->getIntrinsicCost(Intrinsic::ctlz, InitX->getType(), Args) >
1438 TargetTransformInfo::TCC_Basic)
1441 transformLoopToCountable(PH, CntInst, CntPhi, InitX, DefX,
1442 DefX->getDebugLoc(), ZeroCheck,
1443 IsCntPhiUsedOutsideLoop);
1447 /// Recognizes a population count idiom in a non-countable loop.
1449 /// If detected, transforms the relevant code to issue the popcount intrinsic
1450 /// function call, and returns true; otherwise, returns false.
1451 bool LoopIdiomRecognize::recognizePopcount() {
1452 if (TTI->getPopcntSupport(32) != TargetTransformInfo::PSK_FastHardware)
1455 // Counting population are usually conducted by few arithmetic instructions.
1456 // Such instructions can be easily "absorbed" by vacant slots in a
1457 // non-compact loop. Therefore, recognizing popcount idiom only makes sense
1458 // in a compact loop.
1460 // Give up if the loop has multiple blocks or multiple backedges.
1461 if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 1)
1464 BasicBlock *LoopBody = *(CurLoop->block_begin());
1465 if (LoopBody->size() >= 20) {
1466 // The loop is too big, bail out.
1470 // It should have a preheader containing nothing but an unconditional branch.
1471 BasicBlock *PH = CurLoop->getLoopPreheader();
1472 if (!PH || &PH->front() != PH->getTerminator())
1474 auto *EntryBI = dyn_cast<BranchInst>(PH->getTerminator());
1475 if (!EntryBI || EntryBI->isConditional())
1478 // It should have a precondition block where the generated popcount intrinsic
1479 // function can be inserted.
1480 auto *PreCondBB = PH->getSinglePredecessor();
1483 auto *PreCondBI = dyn_cast<BranchInst>(PreCondBB->getTerminator());
1484 if (!PreCondBI || PreCondBI->isUnconditional())
1487 Instruction *CntInst;
1490 if (!detectPopcountIdiom(CurLoop, PreCondBB, CntInst, CntPhi, Val))
1493 transformLoopToPopcount(PreCondBB, CntInst, CntPhi, Val);
1497 static CallInst *createPopcntIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
1498 const DebugLoc &DL) {
1499 Value *Ops[] = {Val};
1500 Type *Tys[] = {Val->getType()};
1502 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
1503 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctpop, Tys);
1504 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
1505 CI->setDebugLoc(DL);
1510 static CallInst *createCTLZIntrinsic(IRBuilder<> &IRBuilder, Value *Val,
1511 const DebugLoc &DL, bool ZeroCheck) {
1512 Value *Ops[] = {Val, ZeroCheck ? IRBuilder.getTrue() : IRBuilder.getFalse()};
1513 Type *Tys[] = {Val->getType()};
1515 Module *M = IRBuilder.GetInsertBlock()->getParent()->getParent();
1516 Value *Func = Intrinsic::getDeclaration(M, Intrinsic::ctlz, Tys);
1517 CallInst *CI = IRBuilder.CreateCall(Func, Ops);
1518 CI->setDebugLoc(DL);
1523 /// Transform the following loop:
1525 /// CntPhi = PHI [Cnt0, CntInst]
1526 /// PhiX = PHI [InitX, DefX]
1527 /// CntInst = CntPhi + 1
1528 /// DefX = PhiX >> 1
1530 /// Br: loop if (DefX != 0)
1531 /// Use(CntPhi) or Use(CntInst)
1534 /// If CntPhi used outside the loop:
1535 /// CountPrev = BitWidth(InitX) - CTLZ(InitX >> 1)
1536 /// Count = CountPrev + 1
1538 /// Count = BitWidth(InitX) - CTLZ(InitX)
1540 /// CntPhi = PHI [Cnt0, CntInst]
1541 /// PhiX = PHI [InitX, DefX]
1542 /// PhiCount = PHI [Count, Dec]
1543 /// CntInst = CntPhi + 1
1544 /// DefX = PhiX >> 1
1545 /// Dec = PhiCount - 1
1547 /// Br: loop if (Dec != 0)
1548 /// Use(CountPrev + Cnt0) // Use(CntPhi)
1550 /// Use(Count + Cnt0) // Use(CntInst)
1552 /// If LOOP_BODY is empty the loop will be deleted.
1553 /// If CntInst and DefX are not used in LOOP_BODY they will be removed.
1554 void LoopIdiomRecognize::transformLoopToCountable(
1555 BasicBlock *Preheader, Instruction *CntInst, PHINode *CntPhi, Value *InitX,
1556 Instruction *DefX, const DebugLoc &DL, bool ZeroCheck,
1557 bool IsCntPhiUsedOutsideLoop) {
1558 BranchInst *PreheaderBr = cast<BranchInst>(Preheader->getTerminator());
1560 // Step 1: Insert the CTLZ instruction at the end of the preheader block
1561 // Count = BitWidth - CTLZ(InitX);
1562 // If there are uses of CntPhi create:
1563 // CountPrev = BitWidth - CTLZ(InitX >> 1);
1564 IRBuilder<> Builder(PreheaderBr);
1565 Builder.SetCurrentDebugLocation(DL);
1566 Value *CTLZ, *Count, *CountPrev, *NewCount, *InitXNext;
1568 if (IsCntPhiUsedOutsideLoop) {
1569 if (DefX->getOpcode() == Instruction::AShr)
1571 Builder.CreateAShr(InitX, ConstantInt::get(InitX->getType(), 1));
1572 else if (DefX->getOpcode() == Instruction::LShr)
1574 Builder.CreateLShr(InitX, ConstantInt::get(InitX->getType(), 1));
1576 llvm_unreachable("Unexpected opcode!");
1579 CTLZ = createCTLZIntrinsic(Builder, InitXNext, DL, ZeroCheck);
1580 Count = Builder.CreateSub(
1581 ConstantInt::get(CTLZ->getType(),
1582 CTLZ->getType()->getIntegerBitWidth()),
1584 if (IsCntPhiUsedOutsideLoop) {
1586 Count = Builder.CreateAdd(
1588 ConstantInt::get(CountPrev->getType(), 1));
1590 if (IsCntPhiUsedOutsideLoop)
1591 NewCount = Builder.CreateZExtOrTrunc(CountPrev,
1592 cast<IntegerType>(CntInst->getType()));
1594 NewCount = Builder.CreateZExtOrTrunc(Count,
1595 cast<IntegerType>(CntInst->getType()));
1597 // If the CTLZ counter's initial value is not zero, insert Add Inst.
1598 Value *CntInitVal = CntPhi->getIncomingValueForBlock(Preheader);
1599 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
1600 if (!InitConst || !InitConst->isZero())
1601 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
1603 // Step 2: Insert new IV and loop condition:
1606 // PhiCount = PHI [Count, Dec]
1608 // Dec = PhiCount - 1
1610 // Br: loop if (Dec != 0)
1611 BasicBlock *Body = *(CurLoop->block_begin());
1612 auto *LbBr = cast<BranchInst>(Body->getTerminator());
1613 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
1614 Type *Ty = Count->getType();
1616 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
1618 Builder.SetInsertPoint(LbCond);
1619 Instruction *TcDec = cast<Instruction>(
1620 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
1621 "tcdec", false, true));
1623 TcPhi->addIncoming(Count, Preheader);
1624 TcPhi->addIncoming(TcDec, Body);
1626 CmpInst::Predicate Pred =
1627 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ;
1628 LbCond->setPredicate(Pred);
1629 LbCond->setOperand(0, TcDec);
1630 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
1632 // Step 3: All the references to the original counter outside
1633 // the loop are replaced with the NewCount -- the value returned from
1634 // __builtin_ctlz(x).
1635 if (IsCntPhiUsedOutsideLoop)
1636 CntPhi->replaceUsesOutsideBlock(NewCount, Body);
1638 CntInst->replaceUsesOutsideBlock(NewCount, Body);
1640 // step 4: Forget the "non-computable" trip-count SCEV associated with the
1641 // loop. The loop would otherwise not be deleted even if it becomes empty.
1642 SE->forgetLoop(CurLoop);
1645 void LoopIdiomRecognize::transformLoopToPopcount(BasicBlock *PreCondBB,
1646 Instruction *CntInst,
1647 PHINode *CntPhi, Value *Var) {
1648 BasicBlock *PreHead = CurLoop->getLoopPreheader();
1649 auto *PreCondBr = cast<BranchInst>(PreCondBB->getTerminator());
1650 const DebugLoc &DL = CntInst->getDebugLoc();
1652 // Assuming before transformation, the loop is following:
1653 // if (x) // the precondition
1654 // do { cnt++; x &= x - 1; } while(x);
1656 // Step 1: Insert the ctpop instruction at the end of the precondition block
1657 IRBuilder<> Builder(PreCondBr);
1658 Value *PopCnt, *PopCntZext, *NewCount, *TripCnt;
1660 PopCnt = createPopcntIntrinsic(Builder, Var, DL);
1661 NewCount = PopCntZext =
1662 Builder.CreateZExtOrTrunc(PopCnt, cast<IntegerType>(CntPhi->getType()));
1664 if (NewCount != PopCnt)
1665 (cast<Instruction>(NewCount))->setDebugLoc(DL);
1667 // TripCnt is exactly the number of iterations the loop has
1670 // If the population counter's initial value is not zero, insert Add Inst.
1671 Value *CntInitVal = CntPhi->getIncomingValueForBlock(PreHead);
1672 ConstantInt *InitConst = dyn_cast<ConstantInt>(CntInitVal);
1673 if (!InitConst || !InitConst->isZero()) {
1674 NewCount = Builder.CreateAdd(NewCount, CntInitVal);
1675 (cast<Instruction>(NewCount))->setDebugLoc(DL);
1679 // Step 2: Replace the precondition from "if (x == 0) goto loop-exit" to
1680 // "if (NewCount == 0) loop-exit". Without this change, the intrinsic
1681 // function would be partial dead code, and downstream passes will drag
1682 // it back from the precondition block to the preheader.
1684 ICmpInst *PreCond = cast<ICmpInst>(PreCondBr->getCondition());
1686 Value *Opnd0 = PopCntZext;
1687 Value *Opnd1 = ConstantInt::get(PopCntZext->getType(), 0);
1688 if (PreCond->getOperand(0) != Var)
1689 std::swap(Opnd0, Opnd1);
1691 ICmpInst *NewPreCond = cast<ICmpInst>(
1692 Builder.CreateICmp(PreCond->getPredicate(), Opnd0, Opnd1));
1693 PreCondBr->setCondition(NewPreCond);
1695 RecursivelyDeleteTriviallyDeadInstructions(PreCond, TLI);
1698 // Step 3: Note that the population count is exactly the trip count of the
1699 // loop in question, which enable us to convert the loop from noncountable
1700 // loop into a countable one. The benefit is twofold:
1702 // - If the loop only counts population, the entire loop becomes dead after
1703 // the transformation. It is a lot easier to prove a countable loop dead
1704 // than to prove a noncountable one. (In some C dialects, an infinite loop
1705 // isn't dead even if it computes nothing useful. In general, DCE needs
1706 // to prove a noncountable loop finite before safely delete it.)
1708 // - If the loop also performs something else, it remains alive.
1709 // Since it is transformed to countable form, it can be aggressively
1710 // optimized by some optimizations which are in general not applicable
1711 // to a noncountable loop.
1713 // After this step, this loop (conceptually) would look like following:
1714 // newcnt = __builtin_ctpop(x);
1717 // do { cnt++; x &= x-1; t--) } while (t > 0);
1718 BasicBlock *Body = *(CurLoop->block_begin());
1720 auto *LbBr = cast<BranchInst>(Body->getTerminator());
1721 ICmpInst *LbCond = cast<ICmpInst>(LbBr->getCondition());
1722 Type *Ty = TripCnt->getType();
1724 PHINode *TcPhi = PHINode::Create(Ty, 2, "tcphi", &Body->front());
1726 Builder.SetInsertPoint(LbCond);
1727 Instruction *TcDec = cast<Instruction>(
1728 Builder.CreateSub(TcPhi, ConstantInt::get(Ty, 1),
1729 "tcdec", false, true));
1731 TcPhi->addIncoming(TripCnt, PreHead);
1732 TcPhi->addIncoming(TcDec, Body);
1734 CmpInst::Predicate Pred =
1735 (LbBr->getSuccessor(0) == Body) ? CmpInst::ICMP_UGT : CmpInst::ICMP_SLE;
1736 LbCond->setPredicate(Pred);
1737 LbCond->setOperand(0, TcDec);
1738 LbCond->setOperand(1, ConstantInt::get(Ty, 0));
1741 // Step 4: All the references to the original population counter outside
1742 // the loop are replaced with the NewCount -- the value returned from
1743 // __builtin_ctpop().
1744 CntInst->replaceUsesOutsideBlock(NewCount, Body);
1746 // step 5: Forget the "non-computable" trip-count SCEV associated with the
1747 // loop. The loop would otherwise not be deleted even if it becomes empty.
1748 SE->forgetLoop(CurLoop);